一、背景 | Background

English

By late 2026, leading labs and enterprise trainers reported that verified synthetic data constituted 50–70% of tokens in major pretraining and fine-tuning mixes — crossing the threshold from “augmentation” to primary training source. Causes were structural: high-quality human web text largely exhausted; licensing battles restricted crawl corpora; domain-specific human data (medicine, law, code internals) remained expensive and gated; meanwhile frontier models + simulators produced synthetic text, code, multimodal pairs, and tool traces at 100× lower marginal cost with improving fidelity.

The concept evolved from naive self-play (model eats own outputs → collapse) to Synthetic Data 2.0: multi-model consensus filtering, executable verification (code runs, proofs check, physics sims validate), provenance graphs, and quality tiers explicitly entering revised scaling laws (see scaling-laws-moe post). Regulators began asking “synthetic %” disclosures for high-risk models.

中文

到 2026 年末，领先实验室与企业训练方披露 经核验合成数据占 major 预训练/微调 mix 的 50–70% token — 从「增广」跨越为 主力训练源。结构性原因：高质量人类网页文本 largely 枯竭；许可诉讼限制 crawl 语料；领域人类数据（医疗、法律、内部代码）昂贵且门禁严；而 前沿模型+仿真器 以 低两个数量级边际成本 产出合成文本、代码、多模态对与工具 trace，保真度持续提升。

概念从 naive 自玩（模型吃自身输出 → collapse）演进为 Synthetic Data 2.0：多模型共识过滤、可执行验证（代码可跑、证明可检、物理仿真可验）、溯源图谱、质量分级 Explicit 进入修正缩放定律。监管开始要求高风险模型披露 「合成占比」。

二、架构 | Architecture

English

Synthetic data factory architecture (2026):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

Seed Sources（种子）
  ├── Licensed human slices（high-trust anchor）
  ├── Public textbooks / papers（structured）
  └── Simulator states（games, CAD, lab logs）

Generators
  ├── Teacher LLM ensemble（diverse architectures）
  ├── Programmatic templaters（grammar-guided）
  ├── Diffusion / video synth for multimodal
  └── Agent trace replay（tool calls + outcomes）

Verification Layer
  ├── Executors（unit tests, sandboxes, formal checkers）
  ├── Critic models（reject hallucination / toxicity）
  ├── Deduplication + near-duplicate purge
  └── Human spot audit（statistical sampling）

Curation & Mixing
  ├── Quality tier labels（T0 anchor human → T3 bulk synth）
  ├── Dynamic mixer（scaling law optimizer）
  └── Provenance metadata per shard

Training Consumption
  └── Pretrain / SFT / RL with tier-weighted sampling

Anti-collapse rules: Minimum 30% T0/T1 human-anchor in frontier mixes; never train exclusively on single-generator outputs; periodic human eval regression on held-out real benchmarks.

中文

2026 合成数据工厂： 种子（许可人类切片、结构化公域、仿真状态）→ 多类生成器（教师 LLM 集成、模板、扩散、Agent trace）→ 验证层（执行器、批评模型、去重、人工抽检）→ 分级混合（T0 人类锚点→T3 bulk 合成）→ 训练消费。

防 collapse 规则： 前沿 mix 至少 30% T0/T1 人类锚点；禁止 单一生成器独占；定期 人类 eval 回归。

三、趋势 | Trends

English

1. Synthetic data marketplaces — buy verified shards by domain (finance QA, ICD-10 traces).
2. Sim-to-text pipelines — Unity/Unreal logs → caption + reasoning datasets.
3. Legal precedents — courts rule on copyright of synthetic-from-copyrighted prompts (jurisdiction-split).
4. Enterprise default — internal fine-tunes use company synthetic from redacted docs + agents.
5. Benchmark shift — “real-world holdout” suites gain prestige over synthetic-friendly benchmarks.
6. Alignment synthetic — preference pairs generated + verified by debate models + human audit sample.

中文

1. 合成数据市场 — 按领域购买 verified shard。
2. Sim-to-text — 游戏/仿真日志 → caption+推理数据集。
3. 法律先例 — 合成是否侵犯 prompt 版权（法域分化）。
4. 企业默认 — 内部微调用脱敏文档+Agent 生成的 公司合成数据。
5. Benchmark 转向 — 「真实世界 holdout」套件更受重视。
6. 对齐合成 — 辩论模型生成 preference + 人工 audit 样本。

四、优缺点 | Pros and Cons

English

Pros: Unlimited scale; domain coverage; privacy (no raw PII in mix); balanced long-tail tasks; reproducible dataset versioning; cost efficiency.

Cons: Model collapse if verification weak; bias amplification from teacher models; legal uncertainty; eval overfitting to synthetic-friendly metrics; anchor drift if human slice too small; trust erosion if undisclosed synthetic %.

中文

优点： 规模无限；领域覆盖；隐私友好；长尾可平衡；版本可复现；成本低。

缺点： 验证弱则 collapse；教师 偏见放大；法律不确定；eval 过拟合 合成友好指标；锚点过小则 漂移；未披露合成占比则 信任侵蚀。

五、应用场景 | Use Cases

六、GitHub 生态 | GitHub Ecosystem

Synthetic provenance: Emerging data-card.json standard in repos documents generator model hash, verifier version, and tier — adopted by FlagOpen ecosystem trainers.

七、深入探讨 | Extended Discussion

English

Synthetic Data 2.0 distinguishes generators from verifiers — often different model families to reduce self-reinforcing bias. Code synthetic pipelines run pytest + mutation testing; math pipelines use SymPy / Lean checkers; medical text passes UMLS consistency + clinician sample review. Provenance graphs link each shard to {generator, verifier, seed_hash, tier} stored beside parquet in HuggingFace-style repos.

Enterprise trainers built internal synthetic factories on redacted Confluence/PDF: Agent extracts facts → generates Q&A → critic rejects unsupported claims → only approved shards enter mix. Legal signed off when no verbatim PII leaves enclave and synthetic does not memorizable-regurgitate source ( tested via membership inference probes).

Regulatory disclosure: EU AI Act annex templates ask synthetic % by tier; US FDA draft guidance on AI medical devices requests data lineage including sim sources. Benchmark gaming fears led REAL-Bench 2026 — holdout human-collected tasks never shown to major generators.

中文

Synthetic Data 2.0 区分 生成器 与 验证器 — 常为不同模型族以防 自我强化偏见。代码合成跑 pytest+变异测试；数学用 SymPy/Lean；医疗文本过 UMLS 一致性+临床样本审查。溯源图 将每 shard 链至 {generator, verifier, seed_hash, tier} 存于 parquet 旁 HF 式 repo。

企业训练方 在脱敏 Confluence/PDF 上建 内部合成工厂：Agent 抽事实→生成 Q&A→批评模型拒无据 claim→仅 approved shard 入 mix。法务 在 无 verbatim PII 出 enclave 且合成 不可 memorizable 复述 源（membership inference 探针测）时放行。

监管披露： EU AI Act 附件模板问 分级合成占比；FDA AI 器械草案要求含 sim 源的 数据 lineage。Benchmark 刷分 担忧催生 REAL-Bench 2026 — 生成器未见过的 holdout 人类任务。

7.1 合成占比与 benchmark 表现 | Synthetic % vs. Benchmark (illustrative)

八、参考链接 | References

• Shumailov et al., “Model collapse” follow-up studies (2025–2026)
• Epoch AI data stock reports
• EU AI Act training data documentation guidance
• 本系列：ai-timeline-2025-synthetic-data, ai-timeline-2026-scaling-laws-moe

Summary | 总结

In 2026, synthetic data is not a cheat code — it is the main fuel, governed by verification tiers, human anchors, and provenance — without which frontier scaling stalls.

2026 年 合成数据非捷径而是主燃料，由验证分级、人类锚点与溯源治理 — 缺失则前沿缩放停滞。

一、背景 | Background

English

Edge AI in 2024–2025 meant many small specialist models (ASR, vision, tiny chat) per device class. In 2026, Edge Universal LLMs (E-LLM) — single general-purpose language–vision–action backbones distilled to 0.5B–8B parameters — shipped across phones, PCs, IoT gateways, and vehicles with unified tokenizer, chat format, and tool API. Apple Intelligence 2, Qualcomm AI Hub universal stacks, MediaTek NeuroPilot LLM, and open Llama-Edge-3B class models demonstrated >GPT-3.5-quality on common tasks at <500ms first-token latency on NPUs.

Drivers included: NPU TOPS doubling (50–100 INT8 TOPS on flagship phones), speculative decoding on-device, KV-cache compression, and cloud-edge hybrid routing that seamlessly escalates hard queries. Privacy regulation and offline-first UX made on-device universal models a product requirement, not a demo.

中文

2024–2025 边缘 AI 意味着每类设备 多个小专用模型（ASR、视觉、微型聊天）。2026 年 边缘通用大模型（E-LLM） — 蒸馏至 0.5B–8B 的 通用语言–视觉–动作骨干 — 跨 手机、PC、IoT 网关、车载 交付，统一 tokenizer、对话格式与工具 API。Apple Intelligence 2、高通 AI Hub、联发科 NeuroPilot LLM、开源 Llama-Edge-3B 级模型在 NPU 上 首 token <500ms** 实现常见任务 **>GPT-3.5 级质量。

驱动因素：NPU TOPS 翻倍、端侧投机 decode、KV 压缩、云边混合路由 无缝升级难 query。隐私法规与 离线优先 UX 使端侧通用模型成为 产品刚需。

二、架构 | Architecture

English

Edge Universal LLM stack:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

Unified Model Core（0.5B–8B, multimodal optional）
  ├── Transformer / hybrid SSM backbone
  ├── Vision encoder（shared across phone/PC scale）
  └── Action / tool head（function calling, IoT schema）

Runtime Layer
  ├── NPU delegate（Core ML, QNN, NNAPI, CANN edge）
  ├── CPU/GPU fallback paths
  ├── Speculative draft model（tiny 100M assistant）
  └── Dynamic quant（INT4/FP8 per layer sensitivity）

System Integration
  ├── OS-level AI session（memory budget, thermal caps）
  ├── Secure enclave for keys + personal adapter
  ├── Federated / local LoRA personalizations
  └── Hybrid router（on-device vs. cloud escalation）

Developer API
  └── Same OpenAI-compatible / MCP surface on all form factors

Cross-device continuity: User starts task on phone; same E-LLM session state (compressed) syncs to PC via E2E encrypted channel for continuation — standardized in 2026 OS vendor SDKs.

中文

E-LLM 栈： 统一模型核心 → 运行时（NPU 委托、投机 draft、动态量化）→ 系统整合（OS AI 会话、安全 enclave、本地 LoRA、混合路由）→ 统一开发者 API。

跨设备连续： 手机发起任务，压缩会话状态 E2E 同步至 PC 续作 — 2026 OS SDK 标准化。

三、趋势 | Trends

English

1. One model SKU per OEM generation — replaces 5–10 tiny models.
2. Personalization without upload — on-device LoRA from usage (differential privacy).
3. Edge–cloud parity tools — same prompt works; router decides execution site.
4. Real-time multimodal — camera + mic streaming into E-LLM at 15–30 FPS effective.
5. Energy-aware inference — OS throttles decode width on low battery.
6. Open weights race — Llama-Edge, Qwen-Edge, Mistral-Edge compete on NPU benchmarks.

中文

1. 每代 OEM 单一模型 SKU 替代 5–10 小模型。
2. 不上传个性化 — 差分隐私端侧 LoRA。
3. 云边 parity 工具 — 同 prompt，路由决定执行位置。
4. 实时多模态 — 相机+麦克风流式输入。
5. 能耗感知推理 — 低电量缩 decode 宽度。
6. 开源权重竞赛 — NPU benchmark 对标。

四、优缺点 | Pros and Cons

English

Pros: Privacy; offline reliability; low marginal inference cost; consistent UX across devices; reduced cloud egress fees; faster perceived latency.

Cons: Quality ceiling vs. cloud frontier models; OTA size (GB-class updates); fragmentation across NPU SDKs despite universal API; thermal throttling on sustained use; security of on-device adapters storing personal data.

中文

优点： 隐私；离线可靠；边际成本低；跨设备 UX 一致；省 cloud egress；感知延迟低。

缺点： 较 cloud frontier 质量上限；OTA 体积 大；NPU SDK 碎片化；长时 温控降频；个人 adapter 安全。

五、应用场景 | Use Cases

六、GitHub 生态 | GitHub Ecosystem

Qualcomm AI Hub and Google AI Edge publish reference E-LLM conversion pipelines linked from community GitHub mirrors.

七、深入探讨 | Extended Discussion

English

Hybrid routing algorithms in 2026 OS stacks classify queries in <50ms using tiny classifier models: on-device if privacy tag=high OR connectivity=offline OR latency SLA <300ms; else cloud escalate with session context bundle (compressed KV + tool state). Users perceive single assistant personality — brand tuning applied consistently via shared system prompt hash across edge and cloud endpoints.

Quantization advances: mixed-precision per layer chosen by sensitivity analysis; INT4 groupwise with outlier channel FP16 bypass; KV-cache INT8 with negligible perplexity delta on 7B models. Speculative decoding pairs 7B main model with 100M draft trained distantly on same tokenizer — acceptance rates 75–85% on chat workloads.

OEM differentiation shifts from parameter count to personalization quality and thermal sustained performance — Geekbench-style “AI endurance” tests measure tokens/sec after 10-minute stress. Enterprise MDM policies gate which cloud endpoints E-LLM may escalate to (data residency).

中文

2026 OS 混合路由 用微型分类器 <50ms 判定：privacy=high 或 offline 或延迟 SLA <300ms 则 端侧；否则 云端升级 并传 压缩 KV+工具状态 会话包。用户感知 单一助手人格 — 云边通过 共享 system prompt hash 一致品牌调优。

量化进展： 敏感度分析 逐层混合精度；INT4 groupwise+outlier 通道 FP16 bypass；KV INT8 对 7B perplexity 影响可忽略。投机 decode 7B 主模型配 100M draft 同 tokenizer 蒸馏 — 聊天 接受率 75–85%。

OEM 差异化 从 参数量 转向 个性化质量 与 温控 sustained 性能 — Geekbench 式 「AI 耐力」 测 10 分钟 stress 后 tokens/sec。企业 MDM 策略 gate E-LLM 可升级的云端点（数据驻留）。

7.1 云边能力分界（2026 典型）| Edge vs. Cloud Split

八、参考链接 | References

• Apple Intelligence technical reports (2026)
• Qualcomm AI Hub universal LLM guides
• ExecuTorch documentation
• 本系列：ai-timeline-2025-edge-llm-npu

Summary | 总结

2026 Edge Universal LLMs unify on-device AI under one backbone, one API, hybrid escalation — general intelligence at the edge becomes default, not a patchwork of micro-models.

2026 边缘通用大模型以 单一骨干、单一 API、混合升级 统一端侧 AI — 边缘通用智能成为默认而非微模型拼盘。

一、背景 | Background

English

AI for Science progressed from static prediction (AlphaFold) and literature mining to closed-loop autonomous experimentation in 2026. Autonomous Science Systems (ASS) coupled LLM planners with robotic lab equipment (liquid handlers, synthesis stations, microscopes, spectrometers) to execute hypothesis → protocol → run → analyze → revise cycles with minimal human intervention.

Breakthrough deployments appeared in materials discovery (battery electrolytes, catalysts), drug lead optimization (automated SAR loops), and synthetic biology (DBTL: Design-Build-Test-Learn). A landmark 2026 Nature-submitted batch reported AI-directed labs completing 100+ experimental iterations per week, versus ~10 for human-only teams on comparable setups. Humans shifted to goal setting, safety approval, and anomaly adjudication.

中文

AI for Science 从静态预测（AlphaFold）与文献挖掘，在 2026 年演进为 闭环自主实验。自主科学系统（ASS） 将 LLM 规划器与 机器人实验设备（移液工作站、合成站、显微镜、谱仪）耦合，以极少人工干预执行 假设→方案→运行→分析→修订 循环。

里程碑部署出现在 材料发现（电池电解液、催化剂）、药物先导优化（自动化 SAR）、合成生物学（DBTL）。2026 年一批 Nature 级投稿报告 AI 主导实验室每周 100+ 实验迭代，可比纯人工 setup 约 10 次。人类转向 目标设定、安全审批与异常裁决。

二、架构 | Architecture

English

Autonomous lab architecture:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

Scientific Goal Layer
  └── Human: target property, constraints, budget

AI Scientist Agent
  ├── Literature / knowledge graph RAG
  ├── Hypothesis generator
  ├── Protocol synthesizer（equipment-aware）
  └── Bayesian / active learning optimizer

Lab OS / Orchestrator
  ├── LIMS integration
  ├── Robotic workcell scheduler（Opentrons, Chemspeed, custom）
  ├── Instrument drivers（HPLC, NMR API, SEM）
  └── Real-time safety interlocks

Analysis Pipeline
  ├── Auto peak picking / structure ID
  ├── Compare to simulation (DFT, MD)
  └── Update surrogate model → next experiment proposal

Human Gate
  └── Approve hazardous / novel chem / budget overrun

Data flywheel: Every run logs structured provenance (reagents, parameters, raw files, embeddings) into a experiment graph training smaller specialist models and improving the planner.

中文

自主实验室架构： 科学目标层 → AI Scientist Agent（文献 RAG、假设、设备感知方案、主动学习）→ Lab OS（LIMS、机器人调度、仪器驱动、安全联锁）→ 分析流水线 → 人工门（危险品/新颖化学/超预算）。

数据飞轮： 每次运行结构化 provenance 写入 实验图谱，训练 specialist 模型并改进规划器。

三、趋势 | Trends

English

1. Cloud labs as a service — submit goals remotely, robots execute 24/7.
2. Multi-lab federation — agents share experiment graphs (privacy-preserving).
3. Regulatory frameworks — FDA/EMA discussion papers on AI-generated protocols.
4. Reproducibility APIs — one-click replay of agent experiment chains.
5. Cost curves — per-experiment cost down 50% vs. 2024 automated partial loops.
6. Education — grad programs in “AI lab stewardship” emerge.

中文

1. 云实验室即服务 — 远程提交目标，机器人 7×24 执行。
2. 多 lab 联邦 — Agent 共享实验图（隐私保护）。
3. 监管框架 — FDA/EMA 讨论 AI 生成方案。
4. 可复现 API — 一键 replay Agent 实验链。
5. 成本曲线 — 单实验成本较 2024 半自动 loop 降约 50%。
6. 教育 — 「AI 实验室 stewardship」研究生项目出现。

四、优缺点 | Pros and Cons

English

Pros: Massive throughput; unbiased exploration of parameter space; 24/7 operation; automatic documentation; faster iteration on materials and molecules.

Cons: Novel hazard discovery (unexpected exotherms); sim-to-lab gap; IP ownership of AI-discovered compounds; equipment downtime cascades; publication ethics — who is author?; reproducibility across lab hardware variants.

中文

优点： 通量大；参数空间探索无偏；7×24；自动文档；材料/分子迭代更快。

缺点： 未知 hazard；sim-to-lab 差距；AI 发现物 IP 归属；设备故障级联；发表伦理；跨硬件 可复现性。

五、应用场景 | Use Cases

六、GitHub 生态 | GitHub Ecosystem

FlagOpen/FlagOS appears in large-scale simulation coupling for materials (DFT throughput on heterogeneous HPC).

七、深入探讨 | Extended Discussion

English

Self-driving labs in 2026 standardize on LabOS middleware — vendor-agnostic layer above LIMS and robots. Protocols compile to device-specific scripts (Opentrons Python, SiLA2 REST) from a single Agent-authored YAML validated against equipment capability schemas. When a spectrometer returns unexpected peaks, the Analysis Agent proposes contamination vs. novel product hypotheses and schedules confirmatory runs automatically.

Safety interlocks are non-negotiable: hard limits on temperature, pressure, and incompatible reagent mixes enforced below LLM layer; human approval for never-before-synthesized SMILES above toxicity score threshold; kill switch physical e-stop linked to orchestrator heartbeat. Insurance underwriters require ASS audit logs for coverage.

Scientific quality: journals pilot AI-assisted methods sections auto-generated from provenance graphs; reviewers demand replay packages (data + code + robot scripts). Negative results logged at scale reduce publication bias — a hidden benefit of autonomous loops.

中文

2026 自动驾驶实验室 标准化 LabOS 中间件 — LIMS 与机器人之上的厂商无关层。方案从 Agent 撰写的 单一 YAML 编译为设备脚本（Opentrons Python、SiLA2 REST），经设备能力 schema 校验。谱仪返回异常峰时 Analysis Agent 提出 污染 vs 新产物 假设并自动排确认实验。

安全联锁 不可妥协：温度/压力/不兼容试剂 硬限 在 LLM 层以下强制；超 toxicity 阈值的新 SMILES 人工批准；物理急停链 orchestrator 心跳。ASS 审计日志 成保险承保要求。

科学质量： 期刊试点从 provenance 图 自动生成 AI 辅助方法节；审稿人要求 replay 包（数据+代码+机器人脚本）。规模化记录 阴性结果 减 发表偏倚 — 自主 loop 的隐性收益。

7.1 吞吐对比 | Throughput Comparison (typical week)

八、参考链接 | References

• Nature / Science AI-for-science special issues (2025–2026)
• Emerald Cloud Lab, Self-Driving Lab consortium papers
• FDA discussion on AI in drug development
• 本系列：ai-timeline-2025-ai-for-science-pipeline

Summary | 总结

2026 autonomous science closes the loop from AI hypothesis to robotic execution — humans govern goals and safety, machines scale experimentation.

2026 自主科学闭合 AI 假设到机器人执行 环路 — 人类治理目标与安全，机器规模化实验。

一、背景 | Background

English

Task Agents — AI systems that complete multi-step business workflows (create ticket, update CRM, schedule meeting, generate report) rather than only answering chat — became embedded in mainstream enterprise software throughout 2026. Industry surveys (IDC, Forrester, domestic equivalents) consistently reported that ~40% of new or major-version enterprise SaaS products shipped with native task agents: Salesforce Agentforce successors, Microsoft 365 Copilot Tasks, ServiceNow AI Agents, SAP Joule workflows, Feishu/钉钉智能助理, and vertical ERP modules.

The penetration threshold crossed when three conditions aligned: reliable tool calling (schema-validated APIs), enterprise identity integration (SSO + RBAC mirroring human roles), and measurable task completion rates (>85% on bounded workflows in pilots). Chat-only copilots were demoted to entry points; task agents became the unit of ROI.

中文

任务型 Agent — 完成 多步业务流程（建工单、更新 CRM、排会、生成报告）而非仅聊天 — 在 2026 年 嵌入主流企业软件。IDC、Forrester 及国内调研一致显示 约 40% 新发或主版本企业 SaaS 自带原生任务 Agent：Salesforce Agentforce 后继、Microsoft 365 Copilot Tasks、ServiceNow AI Agents、SAP Joule、飞书/钉钉智能助理及 vertical ERP 模块。

渗透阈值 crossing 当三者对齐：可靠工具调用（Schema 校验 API）、企业身份集成（SSO+RBAC 镜像人类角色）、可测任务完成率（试点 bounded 工作流 >85%）。纯聊天 Copilot 降级为 入口；任务 Agent 成为 ROI 单位。

二、架构 | Architecture

English

Enterprise Task Agent reference architecture:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

User Intent（natural language or UI trigger）
    ↓
Intent Router
  ├── Q&A → RAG path（read-only）
  └── Task → Agent path（write-capable）

Task Agent Core
  ├── Planner（decompose into tool steps）
  ├── Memory（session + enterprise graph context）
  ├── Tool Registry（OAuth-scoped SaaS APIs）
  └── Validator（pre/post condition checks）

Execution Engine
  ├── Idempotent tool calls + retry
  ├── Transaction boundaries（rollback on partial fail）
  └── Approval gates（>$10k, PII export, admin ops）

Observability
  ├── Task success/failure metrics
  ├── Cost per completed task
  └── Audit log（SOC2 / 等保）

Deployment models: Embedded (agent runs inside vendor cloud); Private tenant (customer VPC with vendor-managed agent runtime); Bring-your-own-model (BYOM) with vendor agent shell.

中文

企业任务 Agent 参考架构： 意图路由（Q&A vs Task）→ Agent 核心（规划、记忆、工具注册、校验）→ 执行引擎（幂等、重试、事务、审批门）→ 可观测（成功率、单任务成本、审计）。

部署模式： 嵌入式；私有租户 VPC；BYOM（自带模型+厂商 Agent shell）。

三、趋势 | Trends

English

1. Agent marketplaces inside SaaS — install pre-built “Expense Reconciliation Agent” like apps.
2. Cross-app orchestration — one agent spans Salesforce + Workday + internal wiki.
3. Role-based agent personas — same LLM, different tool sets per job title.
4. Pricing shift — per completed task + seat hybrid replaces pure seat SaaS for AI tiers.
5. Union of human + agent queues — shared work queues in ticketing systems.
6. Regulatory task allowlists — finance agents cannot execute non-whitelisted tools.

中文

1. SaaS 内 Agent 应用市场。
2. 跨应用编排 — 单 Agent 跨 CRM+HR+wiki。
3. 角色 Agent 人格 — 同 LLM、不同工具集。
4. 定价转变 — 按完成任务数+席位混合。
5. 人机共享队列 — 工单系统统一队列。
6. 合规任务白名单 — 金融 Agent 仅可调白名单工具。

四、优缺点 | Pros and Cons

English

Pros: Quantifiable productivity (tasks/hour); deep ERP/CRM integration; reduced swivel-chair between apps; 24/7 handling of routine workflows; standardized agent SDKs for ISVs.

Cons: Over-automation risk on edge cases; permission sprawl if RBAC misconfigured; vendor concentration (agent tied to SaaS renewal); user trust when silent failures occur; data residency with cross-app agents.

中文

优点： 生产力可量化；深度集成；减少应用间切换；7×24 Routine 流程；ISV 标准 Agent SDK。

缺点： 边界 case 过度自动化；RBAC 误配 权限蔓延；厂商集中；静默失败 信任 问题；跨应用 数据驻留。

五、应用场景 | Use Cases

六、GitHub 生态 | GitHub Ecosystem

Note: Enterprise SaaS agents often wrap closed APIs, but MCP and OpenAPI-to-tool generators on GitHub accelerate custom task agent builds.

七、深入探讨 | Extended Discussion

English

The 40% penetration figure counts major-version releases and new SKUs with native task agents — not legacy products unchanged since 2023. Penetration varies by category: ITSM/CRM ~55%, ERP ~35%, creative tools ~25% (still chat-first). Task completion rate became the North Star metric in earnings calls alongside seat growth.

Technical enablers beyond tool calling: OAuth-on-behalf-of flows letting agents act as delegated user; idempotency keys on every write API preventing duplicate tickets; optimistic UI with rollback when agent fails mid-workflow; shared memory across chat and record pages so agent knows current Opportunity ID without re-prompting.

Workforce impact: roles shifted from data entry to exception handling — humans manage queues flagged confidence < 0.8 or policy_requires_approval. Unions in EU negotiated disclosure when agent touched customer record and right to human redo within SLA.

中文

40% 渗透 统计 主版本新发 SKU 自带任务 Agent — 非 2023 以来未改 legacy 产品。品类差异：ITSM/CRM ~55%，ERP ~35%，创意工具 ~25%（仍 chat 优先）。任务完成率 与席位增长并列 财报 North Star。

工具调用之外 技术使能：OAuth 代表用户 委派 Agent 行动；写 API 幂等键 防重复工单；Agent mid-workflow 失败 乐观 UI 回滚；聊天与记录页 共享记忆 免重复 prompt Opportunity ID。

劳动力影响： 角色从 录单 转向 异常处理 — 人类处理 confidence < 0.8 或 policy_requires_approval 队列。欧盟工会谈判 Agent 触达客户记录须披露 与 SLA 内 要求人工重做权。

7.1 Task Agent vs. Chat Copilot ROI | ROI Comparison

八、参考链接 | References

• Salesforce / Microsoft / ServiceNow 2026 agent product documentation
• IDC “Worldwide Enterprise AI Applications” forecast
• MCP specification: modelcontextprotocol.io
• 本系列：ai-timeline-2024-enterprise-agent

Summary | 总结

By mid-2026, task agents are default infrastructure in enterprise software — not experimental chatbots — with ROI measured in completed workflows under RBAC and audit.

2026 年中 任务 Agent 已是企业软件默认基础设施 — ROI 以 RBAC 与审计下的 完成任务数 衡量。

一、背景 | Background

English

Between 2023 and 2025, enterprises ran hundreds of AI proofs-of-concept but fewer than 30% reached production (Gartner-style estimates cited across industry reports). Failure modes were repetitive: unclear success metrics, missing eval harnesses, no data governance, security review bottlenecks, and custom snowflake architectures that could not be replicated across business units.

In 2026, Standardized AI MVP Deployment emerged as a repeatable playbook — template architectures, checklists, and reference implementations for verticals (banking, manufacturing, retail, healthcare). Cloud vendors and SIs packaged “MVP-in-a-box” stacks: RAG + agent + observability + policy gates + human review UI, deployable in 2–6 weeks with predefined SLAs. The shift moved AI from innovation theater to factory-line delivery.

Consulting firms published fixed-price MVP SKUs ($150k–$400k) with explicit eval thresholds — if golden set accuracy missed target by >5 points, client paid only discovery phase. This outcome-linked pricing aligned vendor incentives with production success for the first time at scale.

中文

2023–2025 年企业开展大量 AI PoC，但 不足 30% 进入生产（多家行业报告援引的 Gartner 类估算）。失败模式高度重复：成功指标不清、缺评估 harness、无数据治理、安全审查瓶颈、不可复制的雪花架构。

2026 年 行业 AI MVP 标准化落地 成为 可复用 playbook — 面向银行、制造、零售、医疗的模板架构、清单与参考实现。云厂商与 SI 打包 「MVP-in-a-box」：RAG + Agent + 可观测 + 策略门 + 人工复核 UI，2–6 周 部署并带预定义 SLA。AI 从 创新表演 转向 流水线交付。

咨询公司发布 固定价 MVP SKU（15–40 万美元）与 explicit eval 阈值 — 若 golden set 准确率未达标 >5 点，客户仅付 discovery 阶段。此 结果挂钩定价 首次规模化对齐厂商激励与生产成功。

二、架构 | Architecture

English

Reference MVP architecture (2026 standard):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

Experience Layer
  ├── Chat / copilot UI（embed or Teams/Slack）
  └── Task-specific forms（structured intake）

Orchestration Layer
  ├── Agent framework（LangGraph / custom）
  ├── Workflow engine（Temporal / cloud step functions）
  └── Human-in-the-loop queues

Intelligence Layer
  ├── Foundation model router（cost/latency/policy）
  ├── RAG pipeline（chunk, embed, retrieve, rerank）
  ├── Fine-tuned vertical adapter（LoRA / full FT）
  └── Eval runner（golden set, regression on every deploy）

Data & Governance Layer
  ├── Vector DB + document ACL sync
  ├── PII scanner / redaction
  ├── Lineage + audit log（immutable）
  └── Synthetic data augmenter（optional）

Platform Layer
  ├── K8s / serverless
  ├── Secrets + KMS
  ├── Observability（traces, costs, quality scores）
  └── CI/CD with safety gates

MVP delivery phases: Week 1 — KPI workshop + data inventory; Week 2–3 — template deploy + golden dataset; Week 4 — UAT + red-team; Week 5–6 — production hardening + runbook.

中文

2026 参考 MVP 架构： 体验层 → 编排层（Agent+工作流+HITL）→ 智能层（模型路由、RAG、垂直 adapter、Eval）→ 数据治理层 → 平台层（K8s、密钥、可观测、带安全门的 CI/CD）。

交付阶段： 第 1 周 KPI 与数据盘点；2–3 周模板部署与 golden set；第 4 周 UAT+红队；5–6 周生产加固与 runbook。

三、趋势 | Trends

English

1. Vertical MVP catalogs — AWS/Azure/阿里云发布行业模板市场。
2. Eval-first sales — vendors demo on customer’s golden set before contract.
3. Composable modules — swap RAG for fine-tune-only MVP via config flags.
4. Regulatory templates — HIPAA/等保 pre-mapped controls in IaC.
5. Internal AI platforms — Fortune 500 “MVP factory” teams ship 1 MVP/month.
6. Post-MVP scale path — standardized promotion checklist to tier-1 SLA.

中文

1. 垂直 MVP 目录 — 云厂商行业模板市场。
2. Eval 优先销售 — 签约前在客户 golden set 上演示。
3. 可组合模块 — 配置切换 RAG/仅微调 MVP。
4. 合规模板 — HIPAA/等保控制预映射进 IaC。
5. 内部 AI 平台 — 财富 500 MVP 工厂 每月交付 1 个。
6. MVP 后扩展路径 — 标准化升级 tier-1 SLA 清单。

四、优缺点 | Pros and Cons

English

Pros: Predictable time/cost; shared learning across BUs; built-in eval and safety; easier executive ROI reporting; faster vendor comparison (same template baseline).

Cons: Template rigidity — edge cases need custom work; false standardization if teams skip governance modules; vendor template lock-in; underfitting unique competitive workflows; maintenance of golden sets often neglected post-launch.

中文

优点： 可预期时间/成本；BU 间经验复用；内置 eval 与安全；ROI 汇报更易；厂商对比基线统一。

缺点： 模板僵化；跳过治理模块的 伪标准化；厂商模板锁定；独特流程 欠拟合；golden set 上线后维护 neglected。

五、应用场景 | Use Cases

六、GitHub 生态 | GitHub Ecosystem

Enterprise pattern: Monorepo with mvp-template/, eval/golden.json, policies/opa/, deployed via ./deploy-mvp.sh — mirrored in this blog’s deploy-to-root.sh philosophy.

七、深入探讨 | Extended Discussion

English

The MVP factory model treats AI delivery like microservices platform teams: central platform owns templates, security baselines, and observability; business units inject domain golden sets and SME reviewers. A typical 6-week MVP breaks down: Week 1 KPI workshop defines task completion rate target (not vanity DAU); Week 2 data ACL sync proves no cross-BU leakage; Week 3–4 template deploy + eval regression green; Week 5 red-team + legal; Week 6 production SLO + runbook handoff to ops.

Vendor selection shifted to eval RFPs: customers supply 200–500 real (redacted) tasks; vendors run on standard template; score = 0.5·accuracy + 0.3·latency + 0.2·cost with minimum safety gate. Snowflake architectures rejected in favor of config-driven vertical packs — swap vertical=banking in Helm values.

Post-MVP promotion requires 30-day production metrics: task success ≥ target, zero P0 safety incidents, cost per task within budget, golden set regression on every release. Failed promotion rolls back to read-only Q&A mode — a pattern that reduced “demo forever” anti-pattern.

中文

MVP 工厂 将 AI 交付类比 微服务平台团队：中央平台拥有模板、安全基线、可观测；业务单元注入 领域 golden set 与 SME 审查者。典型 6 周 MVP：第 1 周 KPI workshop 定 任务完成率目标（非 vanity DAU）；第 2 周数据 ACL 同步证明 无跨 BU 泄漏；3–4 周模板部署+eval 回归绿；第 5 周红队+法务；第 6 周生产 SLO+runbook 移交运维。

厂商选型 转向 eval RFP：客户提供 200–500 真实（脱敏）任务；厂商在标准模板上跑分；得分=0.5·准确+0.3·延迟+0.2·成本 且过最低安全门。拒绝雪花架构， favor 配置驱动 vertical pack — Helm values 改 vertical=banking 即可。

MVP 后升级 需 30 天生产指标：任务成功率达标、零 P0 安全事件、单任务成本在预算内、每次发布 golden 回归。未通过则回退 只读 Q&A — 减少 「永远 demo」 反模式。

7.1 标准 MVP 清单 excerpt | Standard Checklist Excerpt

• Golden set ≥200 tasks with human labels
• OPA policies for every write tool
• PII scanner on ingest pipeline
• Trace + cost dashboard per task type
• Rollback procedure documented (<15 min RTO)

八、参考链接 | References

• Gartner AI productionization surveys (2025–2026)
• McKinsey “Scaling gen AI in the enterprise” playbooks
• Cloud vendor industry MVP documentation
• 本系列：ai-timeline-2024-rag-enterprise

Summary | 总结

2026 industrializes AI delivery: standard MVP stacks + eval gates + governance-by-default turn PoCs into a factory discipline, not artisanal one-offs.

2026 将 AI 交付 工业化：标准 MVP 栈 + 评估门 + 默认治理，使 PoC 成为工厂纪律而非手工孤例。

Deployment, Migration & Operations Guide for Agent Hermes & OpenClaw

最后更新 | Last updated: 2026-06-06

一、部署模式总览 | Deployment Patterns Overview

中文

个人 Agent 常见四种部署拓扑：

flowchart TB    subgraph A["模式 A：本地 Loopback"]        LAP[笔记本 localhost]        LAP --> GW1[Gateway 仅本机]    end    subgraph B["模式 B：VPS + 消息平台"]        VPS[$5 VPS 长驻]        PHONE[手机 Telegram/WhatsApp]        PHONE --> VPS    end    subgraph C["模式 C：分离式 Gateway/执行"]        GWM[Gateway 机 — 仅消息]        EXE[执行机 — Docker/SSH]        GWM -->|SSH| EXE    end    subgraph D["模式 D：Serverless（Hermes）"]        GWH[Hermes Gateway]        MOD[Modal / Daytona]        GWH --> MOD    end

English

Four deployment patterns: local loopback (safest dev), VPS + messaging (most common prod), split gateway/execution (high security), serverless backends (Hermes Modal/Daytona for near-zero idle cost).

二、Hermes 安装 | Hermes Installation

中文

2.1 一键安装

Linux / macOS / WSL2 / Android (Termux)：

1
2

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc   # 或 source ~/.zshrc

Windows 原生（PowerShell）：

1

iex (irm https://hermes-agent.nousresearch.com/install.ps1)

Windows 推荐路径：WSL2 内运行 bash 安装脚本 — 与 Linux 生产环境一致。

Termux（Android）：直接在手机上运行 Agent，适合轻量 Gateway + Telegram Bot。注意电量与后台进程限制。

2.2 安装器做了什么

2.3 安装布局

2.4 初始化

1
2
3
4
5

hermes setup              # 完整配置向导
hermes setup --portal     # 推荐：OAuth + Tool Gateway 一步完成
hermes model              # 选择 Provider 与模型
hermes tools              # 配置 toolsets
hermes gateway setup      # 配置消息平台

hermes setup --portal 覆盖：模型 Provider + web search + image gen + TTS + cloud browser — 最低摩擦无人值守路径。

English

Install via curl script (Linux/macOS/WSL2/Termux) or PowerShell (native Windows; WSL2 preferred). Installer bundles Python, Node, ripgrep, ffmpeg. Run hermes setup --portal for fastest OAuth + Tool Gateway setup.

2.5 非 root / systemd 服务用户

1
2
3
4
5
6
7
8
9

# 管理员一次性（Debian/Ubuntu）
sudo npx playwright install-deps chromium

# 服务用户
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# 或跳过浏览器：bash -s -- --skip-browser

# 确保 PATH 含 ~/.local/bin/hermes
hermes doctor

English

For systemd service accounts: admin installs Playwright system deps; service user runs installer with --skip-browser if headless only. Verify with hermes doctor.

三、OpenClaw 安装 | OpenClaw Installation

中文

1
2
3

npm install -g openclaw@latest
openclaw onboard --install-daemon
openclaw dashboard

工作区默认：~/.openclaw/workspace/（SOUL.md、AGENTS.md 等）

主配置：~/.openclaw/openclaw.json

English

npm install -g openclaw@latest, then openclaw onboard --install-daemon for guided setup and daemon install. Control UI at http://127.0.0.1:18789/. Workspace at ~/.openclaw/workspace/.

四、渠道配置：Telegram 最快路径 | Channel Setup: Telegram Fastest Path

中文

Telegram 是两框架 上手最快 的渠道之一：Bot Token 申请简单、无需企业资质、长轮询即可运行。

4.1 Hermes Telegram

1
2
3
4
5

hermes gateway setup
# 或编辑 ~/.hermes/.env：
# TELEGRAM_BOT_TOKEN=...
# TELEGRAM_ALLOWED_USERS=123456789
hermes gateway start

生产建议：

• 配置 TELEGRAM_ALLOWED_USERS 或启用 DM pairing
• terminal.backend: docker
• 可选 TELEGRAM_HOME_CHANNEL 用于 Cron 投递

4.2 OpenClaw Telegram

1
2
3
4
5
6
7
8
9

{
  channels: {
    telegram: {
      botToken: "...",
      dmPolicy: "pairing",
      groups: { "*": { requireMention: true } },
    },
  },
}

通过 openclaw dashboard 或 onboard 向导配置。

4.3 渠道对比速查

English

Telegram is the fastest channel for both frameworks. Hermes: hermes gateway setup + allowlist/pairing. OpenClaw: channels.telegram in openclaw.json with dmPolicy: pairing.

五、Gateway 系统服务 | Gateway as System Service

中文

5.1 Hermes

1
2
3
4
5

hermes gateway install              # 用户级 systemd/launchd
sudo hermes gateway install --system  # Linux 开机系统服务
hermes gateway start
hermes gateway stop
hermes gateway stop --all           # 更新前停止所有 Profile

PID 文件：~/.hermes/gateway.pid（Profile 作用域）

后台任务并行运行：Cron 调度（60s tick）、会话过期、记忆 flush、Provider 缓存刷新。

5.2 OpenClaw

openclaw onboard --install-daemon 安装 launchd/systemd 服务。

1
2
3
4
5
6
7

{
  gateway: {
    mode: "local",
    bind: "loopback",           // 生产：loopback 或 auth
    auth: { mode: "token", token: "long-random-token" },
  },
}

远程访问：Tailscale 或 SSH 隧道 — 避免直接公网暴露 18789。

English

Hermes: hermes gateway install (user or --system service). OpenClaw: daemon via onboard. Bind loopback or enable auth; use Tailscale/SSH for remote access.

六、hermes claw migrate 迁移 | Migrating from OpenClaw

中文

1

hermes claw migrate

一键从 ~/.openclaw/ 导入：

迁移后仍需：

1
2
3

hermes model              # 确认 Provider
hermes gateway setup      # 验证渠道 Token
hermes doctor             # 健康检查

适用场景：已有龙虾部署、想叠加 Hermes 学习闭环，或社区 HermesClaw 双栈实验。

English

hermes claw migrate imports SOUL, memory, skills, API keys, and messaging config from ~/.openclaw/. Follow with hermes model, hermes gateway setup, and hermes doctor.

七、诊断与审计 | Diagnostics & Auditing

中文

7.1 hermes doctor

1
2

hermes doctor
hermes doctor --ack <id>    # 确认处置供应链告警

检查项包括：

• Python venv 完整性
• 已知妥协包版本（供应链蠕虫等）
• 配置迁移状态
• 安装方式检测（pip/git/Homebrew/Nix）
• 缺失依赖与修复建议

7.2 openclaw security audit

1
2
3

openclaw security audit
openclaw security audit --deep    # 含实时 Gateway 探测
openclaw security audit --fix     # 自动修复常见问题

覆盖：入站访问、工具爆炸半径、网络暴露、文件权限、插件策略漂移。

7.3 对照表

English

hermes doctor for health and supply-chain advisories. openclaw security audit [--deep] [--fix] for OpenClaw hardening. hermes update auto-detects install method.

八、ACP IDE 集成 | ACP IDE Integration

中文

1
2

pip install -e '.[acp]'    # 或在标准安装后
hermes acp

ACP 使用 hermes-acp 精选 toolset：文件、终端、web、memory、skills、delegate_task — 不含 cronjob、messaging delivery。

审批选项：allow_once / allow_session / allow_always / deny

1

hermes acp --setup-browser --yes   # 可选浏览器工具

English

hermes acp for VS Code, Zed (via ACP Registry + uv), JetBrains. Curated toolset for editor workflows. Configure credentials first with hermes model.

九、轨迹导出与 RL 研究 | Trajectories & RL Research

中文

Hermes 提供 Batch Runner、ShareGPT 轨迹导出、Atropos RL 集成与断点续跑，面向工具调用模型微调。OpenClaw 会话存于 sessions/*.jsonl，可手动提取但无内置 RL 管线。

English

Hermes: batch runner, ShareGPT export, Atropos RL. OpenClaw: jsonl transcripts only — no built-in RL pipeline.

十、移动端 Nodes（OpenClaw）| iOS/Android Nodes

中文

OpenClaw 通过 Web Control UI 配对 iOS/Android Nodes（Canvas、相机、语音）。Hermes 无原生 Node，以 Telegram/WhatsApp 等消息平台作「口袋助理」；Android 可用 Termux 自托管。需手机硬件深度集成选 OpenClaw Nodes；纯消息 + Cron 选 Hermes Gateway。

English

OpenClaw nodes: canvas, camera, voice via Control UI pairing. Hermes: messaging platforms or Termux on Android — no native mobile SDK.

十一、分离式 Gateway/执行（SSH 模式）| Split Gateway / Execution

中文

高安全生产拓扑：

1
2
3
4
5

┌─────────────────────┐         SSH          ┌─────────────────────┐
│  Gateway VPS        │ ──────────────────► │  执行 VPS           │
│  - 仅消息 + 配对     │    terminal.backend  │  - Docker 沙箱      │
│  - 无敏感代码仓库    │         : ssh          │  - GPU / 大磁盘     │
└─────────────────────┘                      └─────────────────────┘

Hermes 配置：

1
2
3
4
5
6

terminal:
  backend: ssh
  ssh:
    host: execution.internal
    user: hermes
    key_path: ~/.ssh/hermes_exec

OpenClaw：sandbox 配置 + remote node 模式。

English

Split trust: Gateway host for messaging only; execution host via SSH/Docker sandbox. Hermes: terminal.backend: ssh. OpenClaw: sandbox + remote nodes.

十二、运维清单 | Operations Checklist

中文

English

Routine ops: logs, weekly doctor/audit, monthly updates and key rotation. Production: allowlists, pairing, sandboxed execution, gateway services, and supply-chain checks for both frameworks.

十三、故障排查 | Troubleshooting

中文

English

Common fixes: PATH for hermes, credentials via hermes model/setup --portal, gateway restart, pairing/allowlists for Telegram, config migrate for Hermes, security audit for OpenClaw. Run hermes doctor / hermes status or openclaw security audit --deep for guided diagnosis.

十四、快速命令对照 | Quick Command Reference

十五、延伸阅读 | Further Reading

• Gateway 架构深度解析 — 部署模式与生产清单
• 安全模型深度解析 — audit 与硬化基线
• 模型 Provider 与成本 — Portal 与 Cron 成本
• 插件体系与 MCP — 扩展安装
• Hermes：Installation、Termux、FAQ
• OpenClaw：https://docs.openclaw.ai/

十六、结语 | Conclusion

中文

部署个人 Agent 的「最快路径」是：Hermes 用 curl + setup --portal + Telegram；OpenClaw 用 npm + onboard + Dashboard。生产环境无论选型，都应完成 配对/allowlist、执行沙箱、诊断审计、Gateway 系统服务 四件事。已有龙虾用户可通过 hermes claw migrate 平滑叠加学习闭环；高安全场景采用 SSH 分离 Gateway/执行。运维不是一次性安装 — hermes doctor 与 openclaw security audit 应纳入日常节奏。

English

Fastest paths: Hermes curl + setup --portal + Telegram; OpenClaw npm + onboard + Dashboard. Production requires pairing/allowlists, execution sandboxing, diagnostics, and gateway services. Migrate from OpenClaw with hermes claw migrate. Split gateway/execution via SSH for high security. Treat hermes doctor and security audit as routine ops, not one-time setup.

Plugin Systems & MCP Ecosystem in Agent Hermes & OpenClaw

最后更新 | Last updated: 2026-06-06

一、扩展哲学对比 | Extension Philosophy Comparison

中文

两个框架都将「核心 Agent 引擎」与「可插拔能力」分离，但扩展面不同：

English

Both separate core engines from pluggable capabilities. OpenClaw extends via workspace files, skills, and in-process Gateway plugins plus channel plugins. Hermes adds a Python plugin system with opt-in general plugins, pip distribution, and first-class bidirectional MCP. OpenClaw treats in-process plugins as trusted; Hermes gates arbitrary user plugins behind plugins.enabled.

二、Hermes 插件发现体系 | Hermes Plugin Discovery

中文

flowchart TB    subgraph Sources["发现来源（后者覆盖同名前者）"]        B[bundled plugins/]        U[~/.hermes/plugins/]        P[.hermes/plugins/ 项目级]        PI[pip entry_points]        N[Nix extraPlugins]    end    subgraph Categories["子类别路由"]        G[通用 plugins/ — tools/hooks/commands]        PL[platforms/ — Gateway 渠道]        IG[image_gen/ — 图像后端]        MEM[memory/ — 记忆 Provider]        CE[context_engine/ — 压缩引擎]        MP[model-providers/ — 推理后端]    end    Sources --> PM[PluginManager]    PM --> Categories

2.1 发现来源

同名碰撞时 后加载者覆盖 — 用户插件可替换内置同名 Provider。

2.2 插件类型

English

Discovery order: bundled → user → project (opt-in) → pip → Nix. Subdirectories route to specialized loaders (memory, context engine, model providers, platforms). Later sources override same-name plugins.

2.3 Opt-in 安全模型（plugins.enabled）

1
2
3
4
5
6

plugins:
  enabled:
    - my-tool-plugin
    - disk-cleanup
  disabled:
    - noisy-plugin

1
2
3
4

hermes plugins                    # 交互式 SPACE 切换
hermes plugins enable my-plugin
hermes plugins disable my-plugin
hermes plugins install user/repo --enable   # 安装并启用

不经过 allowlist 的类别（内置基础设施）：

第三方 ~/.hermes/plugins/platforms/ 必须 opt-in。

English

General plugins require explicit plugins.enabled. Bundled platforms/backends and provider plugins bypass the allowlist by design. Third-party platform adapters need opt-in.

2.4 插件能力一览

2.5 Memory Provider 插件

8 种外部记忆后端（Honcho、Mem0、Hindsight、OpenViking 等）通过 plugins/memory/ 发现：

1
2

memory:
  provider: "honcho"    # 空字符串 = 仅内置 MEMORY.md/USER.md

独占模式 — 同时仅一个 active Provider。详见 记忆系统。

2.6 Context Engine 插件

1
2
3

context:
  engine: "compressor"    # 默认内置 ContextCompressor
  engine: "lcm"           # 插件：无损上下文

用户必须显式设置 — 插件引擎不会自动激活。

English

Plugins can register tools, hooks, commands, skills, platforms, providers, and context engines. Memory and context engines are single-select via config.

三、OpenClaw 插件体系 | OpenClaw Plugin System

中文

3.1 进程内插件

OpenClaw 插件在 Gateway 同一 Node.js 进程内运行 — 与 Gateway 共享内存与凭证，视为可信代码。

1
2
3
4
5
6
7
8

{
  plugins: {
    allow: ["matrix-channel", "nostr-bridge"],  // 显式白名单（推荐）
  },
  security: {
    installPolicy: "allowlist",   // 或相关安装策略
  },
}

3.2 Channel 插件

内置渠道：WhatsApp、Telegram、Discord、Slack、Signal、iMessage 等。

插件渠道：Matrix、Nostr、Twitch、Zalo、Feishu 等通过 bundled 或 external channel plugins 扩展。

flowchart LR    GW[Gateway :18789] --> BC[内置渠道]    GW --> CP[Channel Plugins]    CP --> MX[Matrix]    CP --> NO[Nostr]    CP --> TW[Twitch]

3.3 插件 Skills 分发

OpenClaw Skills 以 skills/*/SKILL.md 存在于 workspace，社区通过 ClawHub 等市场分发。插件可附带 Skills 目录 — Skills 与插件 plugins.allow 独立，但同样应限制写入权限。

3.4 npm Shrinkwrap 供应链

发布包使用 npm-shrinkwrap.json 锁定依赖图，配合 openclaw security audit 检测已知妥协版本。对比 Hermes 的 hermes doctor 供应链告警。

English

OpenClaw plugins run in-process — trusted code. Use plugins.allow allowlists and security.installPolicy. Channel plugins extend connectivity. Published deps locked via npm-shrinkwrap; audit via openclaw security audit --deep.

四、MCP：Hermes 作为客户端 | MCP: Hermes as Client

中文

Model Context Protocol 让 Hermes 连接外部工具服务器（GitHub、Linear、数据库、文件系统等），无需为每个集成编写原生工具。

4.1 配置形态

Stdio 本地子进程：

1
2
3
4
5
6

mcp_servers:
  filesystem:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user/projects"]
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: "***"

HTTP 远程端点：

1
2
3
4

mcp_servers:
  linear:
    url: "https://mcp.linear.app/mcp"
    auth: oauth

4.2 工具注册命名

1

mcp_<server_name>_<tool_name>

每个有工具的服务器还创建 runtime toolset：mcp-<server>。

4.3 凭证过滤（Credential Filtering）

Stdio MCP 子进程 不继承完整 shell 环境：

• 仅传递配置中显式 env + 安全基线
• 降低意外泄漏 OPENROUTER_API_KEY 等的风险
• 对比 OpenClaw 进程内插件可访问 Gateway 级凭证

English

Hermes connects to MCP servers via stdio or HTTP. Tools register as mcp_<server>_<tool>. Stdio servers get filtered env — not the full shell — reducing credential leakage.

4.4 per-server 工具过滤

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

mcp_servers:
  github:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-github"]
    tools:
      include: [create_issue, list_issues]
      prompts: false
      resources: false
  stripe:
    url: "https://mcp.stripe.com"
    tools:
      exclude: [delete_customer]
  legacy:
    url: "https://mcp.legacy.internal"
    enabled: false

4.5 目录与 reload-mcp

1
2
3

hermes mcp                  # 交互式目录安装
hermes mcp install n8n      # 安装 Nous 审核条目
hermes mcp configure linear # 重新选择工具 checklist

运行中修改配置：

1

/reload-mcp

服务器也可推送 notifications/tools/list_changed 动态刷新工具列表。

4.6 MCP 目录信任模型

optional-mcps/ 条目经 PR 审核合并。安装会执行 manifest 中的 bootstrap（git clone、pip install、npm install 等）— 安装前阅读 manifest 的 source: 与 install.bootstrap:。

English

Per-server tool filtering via include/exclude. Catalog entries are PR-gated under optional-mcps/. Use /reload-mcp after config changes; servers can push dynamic tool list updates.

4.7 MCP Sampling

MCP 服务器可通过 sampling/createMessage 请求 Hermes 代为推理 — 对不信任服务器设 sampling.enabled: false，并配置 max_rpm / max_tokens_cap 限流。

English

MCP sampling lets servers request LLM inference — disable for untrusted servers; rate limits apply.

五、MCP：Hermes 作为服务端 | MCP: Hermes as Server

中文

hermes mcp serve 将 Hermes 暴露为 MCP 服务器，供 Cursor、Claude Code、VS Code 等客户端调用 消息能力：

1
2
3
4
5
6
7
8

{
  "mcpServers": {
    "hermes": {
      "command": "hermes",
      "args": ["mcp", "serve"]
    }
  }
}

读操作 无需 Gateway 运行；发消息 需要 Gateway 平台适配器在线。

这与 OpenClaw 通过 Gateway WebSocket 统一渠道不同 — Hermes 选择 stdio MCP 桥接 嵌入外部编码 Agent 工作流。

English

hermes mcp serve exposes messaging tools to MCP clients (Cursor, Claude Code). Reads work without gateway; sends require active platform adapters. Bridges external coding agents to Hermes channels.

六、ACP 与 MCP 的关系 | ACP vs MCP

中文

1
2
3

hermes acp              # 编辑器原生 Agent 体验
hermes mcp serve        # 消息桥 MCP 服务
# config.yaml mcp_servers — 扩展 Hermes 工具面

ACP 使用精选 hermes-acp toolset（含 delegate_task），排除 cronjob、messaging delivery 等不适合编辑器 UX 的工具。

English

ACP: Hermes as agent server for IDE UX. MCP: Hermes as messaging bridge (server) or external tool consumer (client). Complementary, not interchangeable.

七、对比矩阵 | Comparison Matrix

中文

English

Matrix: OpenClaw in-process trusted plugins with channel extensions; Hermes opt-in Python plugins with MCP client/server, ACP IDE path, and specialized provider/memory/context plugin loaders.

八、安全最佳实践 | Security Best Practices

中文

OpenClaw

1. plugins.allow 显式白名单 — 不用则等于加载全部发现项
2. openclaw security audit --deep 定期扫描
3. Skills 目录 chmod 限制写入
4. 不信任来源的 channel plugin 不安装
5. 验证 npm shrinkwrap 完整性

Hermes

1. 仅 hermes plugins enable 审查过的通用插件
2. 项目插件保持 HERMES_ENABLE_PROJECT_PLUGINS=false 除非信任仓库
3. MCP tools.include 最小暴露面
4. 不信任 MCP 服务器禁用 sampling
5. Stdio MCP 的 env 仅填必要变量
6. hermes doctor 检查供应链告警
7. hermes mcp 目录安装前阅读 manifest

English

OpenClaw: plugins.allow, security audit, lock down skills dirs. Hermes: opt-in plugins, minimal MCP tool exposure, disable sampling for untrusted servers, filtered stdio env, hermes doctor.

九、典型集成场景 | Typical Integration Scenarios

中文

English

Scenarios: GitHub/Linear via MCP, custom tools via Hermes plugins, Matrix via channel plugins, Cursor→Telegram via mcp serve, VS Code via ACP, Honcho via memory provider.

十、故障排查 | Troubleshooting

中文

English

Hermes: check plugins.enabled, /reload-mcp, hermes mcp login. OpenClaw: check plugins.allow, channel config, security audit.

十一、延伸阅读 | Further Reading

• 安全模型深度解析 — 供应链与 MCP 凭证隔离
• Gateway 架构深度解析 — Channel 与 Platform 插件
• 模型 Provider 与成本 — model-providers 插件
• Hermes：Plugins、MCP、ACP
• OpenClaw 文档：https://docs.openclaw.ai/

十二、结语 | Conclusion

中文

OpenClaw 的扩展栈是 Workspace + Skills + 进程内 Channel 插件，优势在渠道广度与社区生态，安全关键是 plugins.allow 与 shrinkwrap 供应链。Hermes 的扩展栈是 分层 Python 插件（工具/记忆/上下文/Provider）+ 双向 MCP + ACP，优势在可组合的外部工具生态与 opt-in 默认姿态。实践中常组合使用：Hermes mcp_servers 接 GitHub/Linear，OpenClaw channel plugin 接 Matrix，Cursor 通过 hermes mcp serve 桥接消息 — 插件与 MCP 不是二选一，而是分层装配能力边界。

English

OpenClaw extends via workspace, skills, and in-process channel plugins — maximize connectivity with plugins.allow and supply-chain audits. Hermes extends via layered Python plugins and bidirectional MCP/ACP — compose external tools with opt-in safety. In practice, combine MCP servers for SaaS integrations, channel plugins for niche protocols, and hermes mcp serve to bridge coding agents to messaging — plugins and MCP are layers, not either/or.

Multi-Agent Routing & Sub-Agent Delegation in Agent Hermes & OpenClaw

最后更新 | Last updated: 2026-06-06

一、为什么需要多 Agent | Why Multi-Agent?

中文

单一 Gateway 往往要同时服务：

• 个人 vs 工作人格
• 不同聊天渠道（Telegram 私聊 vs 工作群）
• 并行子任务（研究 A/B/C 同时进行）
• 团队共享 Bot vs 个人助理

两个框架都支持「一个控制平面、多个 Agent 脑」，但路由机制与委派模型不同：

English

A single Gateway often serves multiple personas, channels, parallel workstreams, or team vs personal bots. Both frameworks support multiple agent brains on one control plane, but routing and delegation differ: OpenClaw uses sessionKey + bindings + sessions_spawn; Hermes uses structured session keys, profiles, and delegate_task.

二、OpenClaw 多 Agent 路由 | OpenClaw Multi-Agent Routing

中文

2.1 核心概念

flowchart TB    subgraph Inbound["入站消息"]        WA[WhatsApp personal]        WB[WhatsApp biz]        TG[Telegram DM]    end    subgraph GW["OpenClaw Gateway :18789"]        BIND[bindings 确定性匹配]        ROUTE[sessionKey 路由]    end    subgraph Agents["agents.list"]        A1[main — workspace-personal]        A2[work — workspace-work]        A3[family — workspace-family]    end    WA --> BIND    WB --> BIND    TG --> BIND    BIND --> ROUTE    ROUTE --> A1 & A2 & A3

每个 Agent 是完整信任边界：

2.2 agents.list 与 bindings

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

{
  agents: {
    list: [
      {
        id: "main",
        name: "Personal",
        workspace: "~/.openclaw/workspace",
        tools: {
          allow: ["group:fs", "group:sessions", "agents_list"],
        },
        subagents: {
          allowAgents: ["coder", "research"],  // sessions_spawn 目标白名单
        },
      },
      {
        id: "coder",
        workspace: "~/.openclaw/workspace-coder",
        sandbox: { mode: "all", scope: "agent" },
        tools: {
          deny: ["gateway", "cron", "sessions_spawn"],
        },
      },
    ],
  },
  bindings: [
    {
      agentId: "main",
      match: { channel: "whatsapp", accountId: "personal" },
    },
    {
      agentId: "main",
      match: {
        channel: "whatsapp",
        peer: { kind: "group", id: "120363999999999@g.us" },
      },
    },
  ],
}

路由规则：bindings 按 (channel, accountId, peer, guild/team) 确定性匹配，最具体规则优先。

2.3 session.dmScope 与 DM 隔离

1
2
3
4
5
6

{
  session: { dmScope: "per-channel-peer" },
  channels: {
    whatsapp: { dmPolicy: "pairing" },
  },
}

sessionKey 是路由选择器，不是认证令牌。与 dmPolicy: pairing 组合可硬化多用户场景。

English

OpenClaw routes via agents.list (full per-agent workspace, state, sessions, auth) and deterministic bindings. Each agent is an isolated trust boundary. session.dmScope controls DM isolation; sessionKey routes sessions but does not authenticate users.

2.4 sessions_spawn 与 sessions_send

sessions_spawn — 启动后台子代理：

1
2
3
4
5
6

- 会话键：agent:<agentId>:subagent:<uuid>
- deliver: false（结果以内部事件回传）
- 完成后 announce 到请求者聊天渠道
- 默认 maxConcurrent: 8
- 可继承或覆盖 model/thinking
- 默认 maxSpawnDepth: 1（子代理不能再 spawn）

sessions_send — 向另一会话发送消息（跨会话操作，高风险）。

编排者模式（maxSpawnDepth: 2）：

1

Main → Orchestrator sub-agent → Worker sub-sub-agents

深度 1 编排者保留 sessions_spawn、subagents、sessions_list；深度 2 worker 无 session 工具。

2.5 子代理精简 Bootstrap

子代理从目标 Agent 的 workspace 加载 bootstrap 文件（AGENTS.md、TOOLS.md 等），但 不继承 主会话完整历史。可选 cwd 指定子任务工作目录。

这与 Hermes delegate_task「仅 goal+context」哲学类似，但 OpenClaw 仍注入 workspace 级人格与工具指南。

2.6 allowAgents 门禁（常见踩坑）

sessions_spawn 有两层门禁：

1. 工具 allowlist — 必须包含 sessions_spawn（或 group:sessions）
2. 跨 Agent spawn — 调用方 agents.list[].subagents.allowAgents 必须列出目标 agentId

1
2
3
4
5
6

// 正确：allowAgents 与 tools 同级，不在 tools 下
{
  id: "main",
  tools: { allow: ["group:sessions", "agents_list"] },
  subagents: { allowAgents: ["finance"] },  // 或 ["*"]
}

用 agents_list 工具验证可 spawn 的目标列表。

English

sessions_spawn starts background sub-agents with reduced bootstrap from the target workspace. Cross-agent spawning requires subagents.allowAgents on the caller — separate from tools.allow. sessions_send is a high-risk cross-session primitive; deny by default on untrusted surfaces.

2.7 团队 vs 个人 Agent 模式

硬化基线应对不可信面 deny：gateway、cron、sessions_spawn、sessions_send。

English

Patterns: personal multi-persona via bindings, team bots with per-channel-peer + pairing, orchestrator-specialist with sandboxed worker agents. Harden untrusted surfaces by denying control-plane tools.

三、Hermes 多 Agent 与 Profile 隔离 | Hermes Multi-Agent & Profile Isolation

中文

3.1 Session Key 格式

1

agent:main:{platform}:{chat_type}:{chat_id}

示例：agent:main:telegram:private:123456789

禁止手动拼接 — 使用 build_session_key()。详见 Gateway 架构。

3.2 Profile 隔离（hermes -p）

1
2

hermes -p work gateway start
hermes -p personal chat

每个 Profile 拥有独立：

团队 vs 个人：团队 Bot 用 work Profile + 平台 allowlist；个人进化用 default Profile + 学习闭环。

English

Hermes session keys follow agent:main:{platform}:{chat_type}:{chat_id}. Profiles (hermes -p <name>) fully isolate config, sessions, gateway, and token locks — the primary multi-agent pattern on Hermes.

3.3 跨会话镜像与投递

Hermes Gateway 的 delivery.py 支持跨平台投递，但 Cron 投递不镜像进 Gateway 会话历史（避免消息交替违规）。这与 OpenClaw sessions_send 的跨会话写入是不同层面的能力。

English

Cross-platform delivery exists via delivery.py, but cron deliveries are excluded from gateway session history to preserve message ordering invariants.

四、Hermes delegate_task 委派 | Hermes delegate_task Delegation

中文

4.1 设计理念

flowchart LR    PARENT[父 AIAgent] -->|delegate_task| C1[子代理 1]    PARENT -->|delegate_task| C2[子代理 2]    PARENT -->|delegate_task| C3[子代理 3]    C1 --> S1[摘要回注]    C2 --> S2[摘要回注]    C3 --> S3[摘要回注]    S1 & S2 & S3 --> PARENT

• 每个子代理：独立会话、独立终端、可选独立 toolsets
• 中间工具调用 不进入 父上下文 — 仅最终摘要返回
• 同步执行于父轮次内；父中断则子任务取消

1
2
3
4
5
6
7
8
9
10
11
12

delegate_task(tasks=[
    {
        "goal": "Research WebAssembly edge deployments",
        "context": "Focus on Wasmtime, Wasmer, WASI 2025 progress",
        "toolsets": ["web"],
    },
    {
        "goal": "Review src/auth/ for security issues",
        "context": "Project at /home/user/app. Run: pytest tests/auth/ -v",
        "toolsets": ["terminal", "file"],
    },
])

4.2 与 sessions_spawn 对比

4.3 隔离子代理（Isolated Subagents）

子代理 不知道 父对话任何内容。「修复我们刚讨论的 bug」会失败 — 必须在 context 中写明路径、错误信息、约束。

4.4 并行工作流（Parallel Workstreams）

1
2
3
4

delegation:
  max_concurrent_children: 3    # 默认每批 3 并行，可提高到 30+
  max_spawn_depth: 1            # 默认叶子子代理
  orchestrator_enabled: true

编排者子代理（role="orchestrator"）可保留 delegate_task；叶子子代理默认 禁止 delegate_task、clarify、memory、send_message、execute_code。

4.5 成本优化：delegation.provider

1
2
3

delegation:
  provider: openrouter
  model: google/gemini-3-flash-preview

子代理使用廉价模型跑并行研究，父代理用强模型综合 — 常见 质量/成本 平衡点。

English

delegate_task spawns isolated children with separate sessions and toolsets; only summaries return to the parent. Synchronous within the parent turn — interrupted parents cancel children. Pass explicit goal + context; children have zero conversation history. Override delegation.provider/model for cost optimization.

4.6 典型模式速查

English

Patterns: parallel research (web), code review (terminal+file), gather-then-analyze (execute_code then delegate_task).

五、风险矩阵 | Risk Matrix

中文

English

Risk matrix: cross-agent spawn gates (allowAgents), deny sessions_send on untrusted surfaces, use cron for durable work instead of delegation, isolate DMs with per-channel-peer and pairing.

六、OpenClaw 与 Hermes 选型 | When to Use Which

中文

flowchart TD    Q[需要多 Agent？] --> OC{要后台长期子任务<br/>+ 聊天 announce？}    OC -->|是| OC1[OpenClaw sessions_spawn]    OC -->|否| HM{要并行摘要<br/>+ 父上下文隔离？}    HM -->|是| HM1[Hermes delegate_task]    HM -->|否| PR{要完全隔离配置<br/>+ 凭证？}    PR -->|OpenClaw| OC2[agents.list 多 workspace]    PR -->|Hermes| HM2[hermes -p profiles]

English

Use OpenClaw sessions_spawn for background announced sub-runs and multi-workspace routing via bindings. Use Hermes delegate_task for parallel in-turn summaries and hermes -p for full profile isolation.

七、生产配置清单 | Production Checklist

中文

OpenClaw

• session.dmScope: "per-channel-peer"（多用户）
• dmPolicy: "pairing"
• subagents.allowAgents 显式列出可 spawn 目标（避免 ["*"] 除非可信）
• 不可信面 deny sessions_send、gateway、cron
• 专家 Agent 启用 sandbox.mode: "all"
• agents_list 定期审计可 spawn 列表

Hermes

• 团队 Bot 配置平台 allowlist，禁用 GATEWAY_ALLOW_ALL_USERS
• 并行委派设置 delegation.max_concurrent_children 防止 API 风暴
• 子任务用 delegation.provider 指向 Flash 模型
• 持久任务用 cronjob 而非 delegate_task
• 多 Profile 时确认 Token 锁无冲突
• 子代理 context 含绝对路径与验收标准

English

OpenClaw: per-channel-peer, pairing, explicit allowAgents, deny risky tools, sandbox specialists. Hermes: allowlists, cap concurrent children, cheap delegation models, cron for durable work, explicit subagent context.

八、命令对照 | Command Reference

九、延伸阅读 | Further Reading

• Gateway 架构深度解析 — sessionKey、dmScope、Profile 隔离
• 安全模型深度解析 — sessions_spawn deny 基线
• 模型 Provider 与成本 — delegation.provider 成本优化
• OpenClaw：Multi-agent routing、Sub-agents
• Hermes：Delegation、Delegation Patterns

十、结语 | Conclusion

中文

OpenClaw 的多 Agent 哲学是 bindings 路由多个完整 workspace 脑，用 sessions_spawn 做后台 announce 式子任务，适合「一个 Gateway 服务多渠道、多人格、编排者-专家」拓扑。Hermes 的多 Agent 哲学是 Profile 隔离 + 同步 delegate_task 并行，用极简 context 换父上下文清洁，适合「并行研究/审查 + 强模型综合 + 学习闭环沉淀」。二者可经 hermes claw migrate 迁移人格与技能，但路由与委派语义不可直接互换 — 选型应取决于你需要 后台会话 announce 还是 轮内并行摘要。

English

OpenClaw routes multiple full workspace brains via bindings, with background sessions_spawn sub-runs that announce back — ideal for multi-channel, multi-persona, orchestrator-worker topologies. Hermes isolates via profiles and parallel synchronous delegate_task with minimal context — ideal for in-turn research/review with a strong parent synthesizer. Choose based on whether you need background announced sessions or in-turn parallel summaries.

Model Providers & Token Cost Optimization in Agent Hermes & OpenClaw

最后更新 | Last updated: 2026-06-06

一、成本问题的本质 | The Nature of Agent Cost

中文

个人 AI Agent 的运行成本主要来自三类 Token 消耗：

OpenClaw 的模型选择通常绑定在 Gateway 配置或外部 Agent Runtime（Claude Code、Cursor 等），成本优化侧重 工作区文件瘦身 与 工具爆炸半径。Hermes 将 Provider 解析、凭证轮换、fallback、辅助模型、Prompt 缓存、上下文压缩统一纳入 runtime_provider.py 与 AIAgent 循环——适合需要 模型无关 + 长期无人值守 Cron 的场景。

English

Personal agent costs come from three token buckets:

OpenClaw model choice is typically tied to Gateway config or external runtimes; cost control focuses on workspace slimming and tool blast radius. Hermes unifies provider resolution, credential rotation, fallback, auxiliary models, prompt caching, and context compression in runtime_provider.py and the AIAgent loop — ideal for model-agnostic and unattended cron deployments.

二、Hermes Provider 体系（18+）| Hermes Provider Ecosystem (18+)

中文

Hermes 通过 plugins/model-providers/ 插件注册推理后端，用户插件可覆盖同名内置 Provider。核心解析链：

flowchart LR    REQ[用户消息 / Cron / ACP] --> RES[runtime_provider.py]    RES --> POOL[Credential Pool 轮换]    POOL --> MAIN[主模型 API 调用]    MAIN -->|失败| FB[fallback_providers]    MAIN --> AUX[auxiliary.* 侧任务]    AUX --> COMP[compression / vision / approval]

2.1 主模型槽位（Main Model）

配置位于 ~/.hermes/config.yaml 的 model: 段：

1
2
3
4
5

model:
  provider: openrouter
  default: anthropic/claude-opus-4.7
  base_url: ''
  api_mode: chat_completions

English

Hermes registers inference backends via plugins/model-providers/; user plugins override bundled ones. Resolution flow: request → runtime_provider.py → credential pool → main API call → optional fallback_providers → auxiliary tasks.

Main model config lives under model: in config.yaml. Switch via hermes model, hermes setup --portal, dashboard, or /model (session-only or --global).

2.2 三种 API 模式（api_mode）

Fallback 激活时会按目标 Provider 就地切换 api_mode：Codex → codex_responses，Anthropic → anthropic_messages，其余 → chat_completions。

English

Three API modes: chat_completions (default), codex_responses (OpenAI Codex), anthropic_messages (native Anthropic). Fallback swaps api_mode in-place when activating a backup provider.

2.3 Nous Portal 与 Tool Gateway

hermes setup --portal 是最低摩擦路径：

• 300+ 模型 单一 OAuth 订阅
• Tool Gateway 捆绑：web search、image generation、TTS、cloud browser
• OAuth 自动刷新，适合 Cron 无人值守
• Portal 订阅者对按 Token 计费的 Provider 享 10% 折扣

1
2

hermes setup --portal    # 登录 + 设置 Nous Provider + 启用 Tool Gateway
hermes portal info       # 查看已接入能力

对比单独配置 OPENROUTER_API_KEY + 各工具 API Key，Portal 显著降低 密钥管理成本 与 辅助服务账单碎片度。

English

hermes setup --portal covers 300+ models plus Tool Gateway (search, images, TTS, browser) under one OAuth — ideal for unattended cron with automatic token refresh. Portal subscribers get 10% off token-billed providers.

2.4 OpenRouter 与自定义端点

Hermes 严格隔离 API Key 与 base URL：

• OPENROUTER_API_KEY 仅发往 openrouter.ai 端点
• OPENAI_API_KEY 用于自定义 OpenAI 兼容端点及回退
• provider: custom + custom_providers 列表支持 LM Studio、Together、本地 vLLM 等

避免「配置了 OpenRouter 却把 OpenAI Key 泄漏到自定义 localhost」的常见踩坑。

English

API keys are scoped to their base URLs. OPENROUTER_API_KEY never leaks to custom endpoints; provider: custom supports local and third-party OpenAI-compatible servers.

三、凭证池轮换（Credential Pool）| Credential Pool Rotation

中文

凭证池处理 同 Provider 多 Key 轮换；fallback_providers 处理 跨 Provider 故障转移。执行顺序：先池，后 fallback。

1
2
3
4
5

请求 → 从池选 Key（fill_first / round_robin / least_used / random）
     → 429？先重试一次，再轮换下一 Key（冷却 1h）
     → 402 账单/配额？立即轮换（冷却 24h）
     → 401？尝试 OAuth 刷新，失败则轮换
     → 池耗尽 → 激活 fallback_providers

3.1 快速配置

1
2
3

hermes auth add openrouter --api-key sk-or-v1-second-key
hermes auth add anthropic --type oauth          # Claude Max OAuth
hermes auth list                                # ← 标记当前选中凭证

1
2
3

credential_pool_strategies:
  openrouter: round_robin
  anthropic: least_used

3.2 与 Gateway 并发

凭证池使用线程锁保护 select() / mark_exhausted_and_rotate()，多 Telegram/Discord 会话并发时安全。子代理通过 delegate_task _spawn 时 继承父代理凭证池，同 Provider 子任务可共享轮换能力。

English

Credential pools rotate multiple keys for the same provider before cross-provider fallback kicks in. Strategies: fill_first, round_robin, least_used, random. Thread-safe for concurrent gateway sessions; subagents inherit the parent’s pool.

四、主模型 Fallback 链 | Primary Model Fallback Chain

中文

1
2
3
4
5

fallback_providers:
  - provider: openrouter
    model: anthropic/claude-sonnet-4
  - provider: nous
    model: nous-hermes-3

sequenceDiagram    participant U as 用户消息    participant A as AIAgent    participant P as 主 Provider    participant F as fallback_providers    U->>A: 新轮次开始    A->>P: 调用主模型    P-->>A: 429 / 503    A->>F: _try_activate_fallback()    F-->>A: 切换 provider+client+api_mode    A->>F: 继续本轮回话    Note over A: 下一轮消息重新尝试主模型

4.1 Fallback 覆盖范围

English

fallback_providers is an ordered list tried on primary failure. Per-turn scope: each new user message retries the primary first; at most one fallback activation per turn. Cron and subagents inherit the chain; auxiliary tasks use their own routing.

五、辅助模型与成本杠杆 | Auxiliary Models & Cost Levers

中文

Hermes 将侧任务从主模型剥离，共 8 个辅助槽位：

1
2
3
4
5
6
7
8
9
10

auxiliary:
  compression:
    provider: openrouter
    model: google/gemini-3-flash-preview
  approval:
    provider: openrouter
    model: anthropic/claude-haiku-4-5
  title_generation:
    provider: openrouter
    model: google/gemini-3-flash-preview

provider: auto 表示使用主模型 — 对 Compression / Approval 通常是 浪费。

English

Eight auxiliary slots offload side jobs from the main model. Override compression and approval with fast/cheap models — using Opus for summarization wastes reasoning tokens. provider: auto means “use main model.”

5.1 Smart Approval 的辅助 LLM 成本

approval_mode: smart 时，每条待审批命令会调用 auxiliary.approval 做风险分类：

成本建议：将 auxiliary.approval 指向 haiku / flash / gpt-5-mini；切勿用 Opus 做审批评分。容器后端（Docker/Modal）跳过审批检查 — 容器即边界。

English

approval_mode: smart routes each dangerous-command candidate through auxiliary.approval. Point it at haiku/flash/mini models — never Opus. Container backends skip approval checks entirely.

5.2 辅助模型容量错误 Fallback

显式配置 auxiliary.vision.provider: glm 等时，若遇 402/日配额耗尽/连接失败，Hermes 按层回退：

1. 配置的 aux Provider
2. auxiliary.*.fallback_chain（可选）
3. 主代理 Provider + 模型（安全网）
4. 全部失败 → WARNING 日志 + 抛出原错误

瞬时 429（Retry-After）不触发此阶梯，尊重显式 Provider 选择。

English

Explicit auxiliary providers fall back through optional fallback_chain, then the main agent model, on capacity errors (402, daily quota, connection failure) — not transient 429s.

六、上下文压缩（ContextCompressor）| Context Compression

中文

Hermes 采用 双层压缩，防止长会话 Token 爆炸：

flowchart TB    MSG[新消息到达] --> HY[Gateway Session Hygiene 85%]    HY --> AG[Agent ContextCompressor 50%]    AG --> P1[Phase1: 剪枝旧 tool 输出]    P1 --> P2[Phase2: 划定 head/tail 边界]    P2 --> P3[Phase3: 辅助 LLM 结构化摘要]    P3 --> P4[Phase4: 重组消息列表]

1
2
3
4
5
6
7
8
9
10

compression:
  enabled: true
  threshold: 0.50
  target_ratio: 0.20
  protect_last_n: 20

auxiliary:
  compression:
    provider: openrouter
    model: google/gemini-3-flash-preview

关键警告：摘要模型的上下文窗口必须 ≥ 主模型。否则中间段无法一次送入摘要 API，压缩退化为 无摘要丢弃 — 最常见的质量劣化原因。

压缩触发 会话分裂（parent_session_id 链），详见 记忆系统。

English

Dual compression: gateway hygiene at 85% (safety net), agent ContextCompressor at 50% (default). Four phases: prune old tool output, bound head/tail, auxiliary LLM structured summary, reassemble. Summary model context must be ≥ main model or middle turns are dropped without summary.

6.1 可插拔 Context Engine

1
2
3

context:
  engine: "compressor"    # 默认有损摘要
  engine: "lcm"           # 插件：无损上下文管理

插件需用户显式设置 context.engine — 默认 "compressor" 始终使用内置实现。

English

Plugins can replace the context engine via context.engine (e.g., lossless lcm). User must opt in explicitly.

七、Anthropic Prompt Caching | Anthropic Prompt Caching

中文

对 Claude 模型，Hermes 自动启用 cache_control（agent/prompt_caching.py），多轮对话输入成本可降约 75%。

策略 system_and_3（Anthropic 最多 4 个断点）：

1
2

断点 1: 系统提示词（跨轮稳定）
断点 2-4: 倒数第 3/2/1 条非 system 消息（滚动窗口）

1
2

prompt_caching:
  cache_ttl: "5m"

启用条件：Claude 模型名 + Provider 支持 cache_control（原生 Anthropic 或 OpenRouter）。

English

Anthropic prompt caching via system_and_3 strategy: system prompt plus rolling 3-message window. ~75% input cost reduction on multi-turn Claude conversations. Preserve prompt stability; compression appends a note only on first compaction.

八、Cron 成本治理 | Cron Cost Governance

中文

无人值守 Cron 是 Token 成本 放大器。Hermes 提供多层节制：

1
2
3
4
5
6
7
8

cronjob(
    action="create",
    schedule="every sunday 9am",
    enabled_toolsets=["web", "file"],   # 不带 terminal/browser/delegation
    provider="openrouter",
    model="google/gemini-3-flash-preview",
    prompt="Summarize this week's AI news...",
)

反面教材：默认携带 moa、browser、delegation 的 Cron 在每次 LLM 调用中注入大量工具 schema — 对小任务极其浪费。

English

Cron amplifies token cost. Control via enabled_toolsets, platform defaults in hermes tools, no_agent script-only jobs, wakeAgent: false gates, per-job cheap models, and inherited fallback/credential pools. Avoid bloated toolsets on simple scheduled tasks.

九、OpenClaw 模型与成本 | OpenClaw Models & Cost

中文

OpenClaw 模型由 Gateway Runtime 或外部编码 Agent（Claude Code、Cursor）配置，无 Hermes 式 18+ Provider 抽象。成本杠杆：工作区瘦身、tools.profile: messaging、子代理 sessions_spawn 隔离长任务、openclaw security audit 收紧工具面。云账单常见 $10–150+/月；Hermes 对辅助模型、压缩、缓存的可编程控制更细。

English

OpenClaw lacks Hermes-style multi-provider runtime. Cost levers: slim workspaces, tight tool profiles, sessions_spawn isolation, security audit. Cloud bills commonly $10–150+/month; Hermes offers finer aux/compression/caching control.

十、模型选择卫生（Hygiene）| Model Selection Hygiene

中文

English

Hygiene checklist: cheap models for aux tasks, dedicated cron models, stable system prompts for cache hits, credential pools for rate limits, dashboard analytics for monitoring.

十一、成本优化决策树 | Cost Optimization Decision Tree

flowchart TD    START[账单过高？] --> Q1{主模型是否过强？}    Q1 -->|是| A1[降级主模型 / 按任务选模型]    Q1 -->|否| Q2{辅助任务用主模型？}    Q2 -->|是| A2[配置 auxiliary.compression 等 Flash 模型]    Q2 -->|否| Q3{Cron 工具过多？}    Q3 -->|是| A3[enabled_toolsets 精简]    Q3 -->|否| Q4{长会话上下文膨胀？}    Q4 -->|是| A4[调低 compression.threshold / 检查摘要模型窗口]    Q4 -->|否| Q5{Claude 多轮对话？}    Q5 -->|是| A5[确认 prompt caching 已启用]    Q5 -->|否| A6[凭证池 + fallback 防失败重试浪费]

十二、配置速查 | Configuration Quick Reference

中文

English

Quick ref: hermes setup --portal, hermes model, hermes fallback, hermes auth, /model, compression.*, auxiliary.*, enabled_toolsets, prompt_caching.cache_ttl.

十三、延伸阅读 | Further Reading

• 记忆系统深度解析 — ContextCompressor 与会话分裂
• Gateway 架构深度解析 — Gateway 85% 卫生压缩
• 安全模型深度解析 — smart approval 与 Tirith
• Hermes 官方：Configuring Models、Fallback Providers、Context Compression

十四、结语 | Conclusion

中文

Hermes 将 Provider 解析、凭证池、fallback、辅助模型、双层压缩、Anthropic 缓存 串成可配置的成本治理体系；OpenClaw 则通过 工作区瘦身、工具 profile、子代理隔离 控制爆炸半径。实践中的最高 ROI 动作通常是：为 Compression / Title / Approval 配置 Flash 模型、为 Cron 设置 enabled_toolsets、启用凭证池与 fallback 避免失败重试、在 Claude 长会话中依赖 Prompt Caching。模型无关不等于成本无关 — 侧任务与工具 schema 才是隐形大户。

English

Hermes offers a configurable cost stack: providers, credential pools, fallback, auxiliary models, dual compression, and Anthropic caching. OpenClaw leans on workspace slimming, tool profiles, and sub-agent isolation. Highest-ROI moves: flash models for aux tasks, enabled_toolsets for cron, pools + fallback for resilience, prompt caching for long Claude sessions. Model-agnostic doesn’t mean cost-agnostic — auxiliary calls and tool schemas are the hidden spend.

Agent Hermes & OpenClaw: Automation Scheduling and Proactive Monitoring — A Deep Dive

最后更新 | Last updated: 2026-06-06

一、自动化能力概览 | Automation Capability Overview

中文

个人 Agent 的「主动性」取决于能否在无人值守时执行任务。两个框架提供互补机制：

English

Proactivity depends on unattended execution. Both frameworks offer complementary automation:

二、Hermes cronjob 工具全解析 | Hermes cronjob Tool Deep Dive

中文

Hermes 将定时任务管理收敛为单一 cronjob 工具（action 风格），CLI、Gateway、自然语言对话共用同一 API。

2.1 支持的操作

1
2
3
4
5
6
7
8

cronjob(
    action="create",
    name="morning-digest",
    schedule="0 9 * * *",
    skills=["blogwatcher"],
    prompt="Check configured feeds and summarize anything new.",
    deliver="telegram",
)

2.2 调度格式

重复行为：

2.3 Skill-Backed Cron

任务可附加零个、一个或多个技能，执行时按顺序注入：

1
2
3
4
5
6
7

cronjob(
    action="create",
    skills=["blogwatcher", "maps"],
    prompt="Combine new feed items with nearby events into one brief.",
    schedule="every 6h",
    name="local-brief",
)

技能内容作为上下文注入，prompt 仅承载任务指令——避免在 cron prompt 中粘贴完整技能正文。

English

Hermes unifies scheduling in the cronjob tool with actions: create, list, update, pause, resume, run, remove. Schedule formats: relative delays, intervals (every N), cron expressions, natural language, ISO timestamps. Attach zero or more skills loaded in order at execution. Prompt carries task instruction only.

三、workdir 与 profile 钉扎 | workdir & profile Pinning

中文

Cron 任务默认脱离任何代码库运行——不加载 AGENTS.md、.cursorrules，终端/文件工具使用 Gateway 启动目录。

3.1 workdir 钉扎

1
2
3
4
5
6

cronjob(
    action="create",
    schedule="every 1d at 09:00",
    workdir="/home/me/projects/acme",
    prompt="Audit open PRs, summarize CI health, and post to #eng",
)

当 workdir 设置时：

• 注入该目录的 AGENTS.md、.cursorrules（与交互式 CLI 相同发现顺序）
• terminal、read_file、patch、execute_code 使用该目录为 cwd
• 路径必须是存在的绝对路径
• workdir="" 可清除钉扎

序列化约束：带 workdir 的任务在 scheduler tick 上串行执行（进程全局 cwd 状态）。

3.2 profile 钉扎

1
2
3
4
5
6

cronjob(
    action="create",
    schedule="every 1d at 03:00",
    profile="night-ops",
    prompt="Tail the security log and flag anomalies",
)

调度器临时切换 HERMES_HOME 到目标 profile，加载其 .env + config.yaml。带 profile 的任务同样串行执行（HERMES_HOME 是进程全局状态）。

English

Cron jobs default to detached execution without project context. workdir pins AGENTS.md/.cursorrules injection and tool cwd to an absolute project path — serial execution due to global cwd. profile pins HERMES_HOME/config for the run — also serial due to global profile switch.

四、投递选项与静默模式 | Delivery Options & Silent Mode

中文

4.1 deliver 参数

最终响应自动投递，无需 prompt 内 send_message。

4.2 响应包装

默认包装 header/footer 标明来源为定时任务。设 cron.wrap_response: false 可输出原始内容。

4.3 [SILENT] 静默抑制

若 Agent 最终响应以 [SILENT] 开头，抑制外发投递，输出仍保存到 ~/.hermes/cron/output/ 供审计。

1
2

Check if nginx is running. If healthy, respond with only [SILENT].
Otherwise, report the issue.

仅成功运行可静默；失败任务始终投递。

English

deliver routes output to origin, local files, specific platforms, or all fan-out. Final agent response auto-delivers without send_message. [SILENT] prefix suppresses outbound delivery on success while saving locally. Failed jobs always deliver. cron.wrap_response: false removes the wrapper header/footer.

五、no-agent 模式与 wakeAgent 门控 | no-agent Mode & wakeAgent Gate

中文

5.1 no-agent 模式（零 Token 看门狗）

1
2
3
4
5

hermes cron create "every 5m" \
  --no-agent \
  --script memory-watchdog.sh \
  --deliver telegram \
  --name "memory-watchdog"

.sh/.bash 用 /bin/bash；其他用当前 Python 解释器。

5.2 wakeAgent 门控（LLM 任务的 $0 预检）

带 script= 的 LLM 任务，预检脚本末行可输出 JSON 决定是否唤醒 Agent：

1

{"wakeAgent": false}

1

{"wakeAgent": true, "context": {"new_issues": 3}}

典型配方：

flowchart TD    T[Scheduler Tick] --> S{有 script?}    S -->|否| A[直接运行 Agent]    S -->|是| R[运行预检脚本]    R --> W{末行 wakeAgent?}    W -->|false| Z[静默跳过 — $0]    W -->|true/省略| A    A --> D[投递结果]

English

no_agent=True: script-only, zero LLM tokens, stdout delivered verbatim, empty stdout = silent tick. wakeAgent gate: pre-check script emits {"wakeAgent": false} on last stdout line to skip the agent call for that tick — useful for 1-5 min polls that only need the LLM when state changed. Optional context object passes structured data to the agent.

六、context_from 流水线 | context_from Pipelines

中文

Cron 任务在隔离会话中运行，无上次执行记忆。context_from 自动将上游任务最近输出前置到当前 prompt：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

# 阶段 1：采集
cronjob(action="create", name="ai-news-fetch",
        schedule="0 7 * * *",
        prompt="Fetch top 10 AI stories from HN, save to raw.md")

# 阶段 2：筛选（读取阶段 1 最近输出）
cronjob(action="create", name="ai-news-triage",
        schedule="30 7 * * *",
        context_from="ai-news-fetch",
        prompt="Score each story 1-10, output top 5 to ranked.md")

# 阶段 3：发布
cronjob(action="create", name="ai-news-brief",
        schedule="0 8 * * *",
        context_from="ai-news-triage",
        prompt="Write 3 tweet drafts and deliver to telegram:7976161601")

输出从 ~/.hermes/cron/output/{job_id}/*.md 读取，按列表顺序拼接。读取最近已完成输出——不等待同 tick 仍在运行的上游任务。

English

context_from prepends upstream jobs’ most recent completed output from ~/.hermes/cron/output/{job_id}/ to the current prompt. Accepts single job ID/name or list for fan-in. Enables collect → filter → deliver pipelines without glue code or databases.

七、Gateway 调度器 internals | Gateway Scheduler Internals

中文

sequenceDiagram    participant T as 60s Ticker    participant L as .tick.lock    participant J as jobs.json    participant A as AIAgent    participant D as Delivery    T->>L: 获取文件锁    T->>J: 加载任务    T->>T: 筛选 due jobs (next_run <= now)    loop 每个 due job        T->>A: 创建全新会话（无历史）        opt 附加 skills        T->>A: 注入 skills + prompt + context_from        opt script 预检        T->>A: wakeAgent 门控        A->>D: 完成 → 投递        T->>J: 更新 run_count, next_run    end    T->>L: 释放锁

存储：jobs.json（原子写）、cron/output/{job_id}/。Gateway 每 60s tick，.tick.lock 防重叠；Cron 会话禁用 cronjob toolset。enabled_toolsets 可 per-job 收紧工具 schema。

English

60s tick, file lock, fresh sessions, cron toolset disabled in cron runs, enabled_toolsets for cost control, fallback provider inheritance.

八、OpenClaw HEARTBEAT 主动巡检 | OpenClaw HEARTBEAT Proactive Monitoring

中文

OpenClaw 的主动性主要通过 Gateway heartbeat 实现——周期性 Agent turn，默认 30 分钟（Anthropic OAuth 检测时为 1 小时）。

8.1 配置

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

{
  heartbeat: {
    every: "30m",           // "0m" 禁用
    target: "last",         // "none" | "last" | "slack" | "telegram" ...
    activeHours: {
      start: "09:00",
      end: "22:00",
      timezone: "America/New_York",
    },
    schedule: [             // 可选：时段差异化间隔
      { start: "08:00", end: "18:00", every: "15m" },
      { start: "23:00", end: "08:00", every: "2h" },
    ],
  },
}

8.2 HEARTBEAT.md 清单

工作区中的 HEARTBEAT.md 是巡检 checklist——短小、稳定、适合每 30 分钟考虑：

1
2
3
4
5

# Heartbeat Checklist

- Scan inbox for urgent emails (last 30 min)
- Check calendar for meetings in next 2 hours
- Verify production health endpoint returns 200

tasks: 结构化块（任务级独立间隔）：

1
2
3
4
5
6
7

tasks:
  - name: inbox-triage
    interval: 30m
    prompt: Check for urgent emails.
  - name: calendar-scan
    interval: 2h
    prompt: Check for upcoming meetings.

8.3 HEARTBEAT_OK 静默

一切正常时回复 HEARTBEAT_OK — 静默确认，不外发；有异常才 alert 到 target。

English

OpenClaw heartbeat: 30m cadence (1h for Anthropic OAuth), HEARTBEAT.md checklist, optional tasks: per-interval checks, HEARTBEAT_OK silent ack.

九、OpenClaw cron 工具与风险 | OpenClaw cron Tool & Risks

中文

OpenClaw 提供 cron 工具让 Agent 创建持久定时任务——属于高风险控制面工具：

硬化：tools.deny: ["cron", "gateway", "sessions_spawn"]；不可信面必须 deny；openclaw security audit 定期审查。Hermes 对等：Cron 内禁用 cronjob + Prompt 扫描。

English

OpenClaw cron is high-risk persistent scheduling — deny on untrusted surfaces, minimal profiles, security audit. Hermes: cron toolset disabled in cron runs + prompt scanning.

十、安全：Cron Prompt 扫描 | Security: Cron Prompt Scanning

中文

Hermes 在创建/更新时扫描 prompt：注入、凭证外泄、不可见 Unicode、SSH 后门等——阻断则拒绝创建并返回明确错误。运行时：cron_mode: deny（无人值守推荐）、enabled_toolsets 限制、脚本限于 ~/.hermes/scripts/、script_timeout_seconds 默认 120s。

English

Create/update prompt scanning blocks injection and exfiltration. Runtime: cron_mode: deny, limited toolsets, script sandbox, configurable timeout.

十一、选型速查 | Selection Quick Reference

十二、最佳实践与命令速查 | Best Practices & Quick Commands

中文

Hermes Cron

1. 自包含 prompt：Cron 会话无历史，须写清一切必要细节
2. 技能而非长 prompt：复用工作流用 skills=[...] 附加
3. 收敛 toolsets：enabled_toolsets=["web","file"] 控制成本
4. 健康检查用 [SILENT]：正常时静默，异常才打扰
5. 流水线用 context_from：避免硬编码文件路径在 prompt 中
6. 生产用 Nous Portal OAuth：无人值守避免 API key 过期

OpenClaw Heartbeat

1. 保持 HEARTBEAT.md 短小：<20 行 checklist
2. tasks: 分间隔：不同检查项用不同 interval
3. activeHours 限制：避免深夜无意义 Token 消耗
4. target: “last”：有异常时发到最近活跃渠道
5. deny cron 给聊天 Agent：heartbeat 与 cron 职责分离

English

Hermes: self-contained prompts, skills attachment, limited toolsets, [SILENT] health checks, context_from pipelines, Nous Portal OAuth. Commands: /cron add|list|pause|resume|run, hermes gateway install.

OpenClaw: short HEARTBEAT.md, per-task intervals, activeHours, target last channel, deny cron on chat agents.

十三、延伸阅读 | Further Reading

• 工具链与执行环境
• 工作区文件与 Prompt 组装
• 安全模型深度解析
• Hermes Cron 文档
• OpenClaw Heartbeat 文档

十四、结语 | Conclusion

中文

自动化调度与主动巡检让个人 Agent 从「被动应答」进化为「持续值守」。OpenClaw 以 HEARTBEAT.md + 30 分钟 heartbeat + HEARTBEAT_OK 静默 提供轻量、内置的主动性，适合多渠道助理的日常巡检。Hermes 以 60 秒调度器、cronjob 统一 API、context_from 流水线、no-agent 零 Token 看门狗和 wakeAgent 门控 提供工业级无人值守能力，适合复杂流水线和成本敏感的高频轮询。安全配置的核心原则一致：不可信面 deny 调度工具，自包含 prompt，最小 toolset，失败必告警。

English

Automation and proactive monitoring evolve personal agents from reactive to always-on. OpenClaw offers lightweight proactivity via HEARTBEAT.md, 30m heartbeat, and HEARTBEAT_OK silent ack — ideal for multi-channel daily checks. Hermes offers industrial unattended capability via 60s scheduler, unified cronjob API, context_from pipelines, no-agent zero-token watchdogs, and wakeAgent gates — ideal for complex pipelines and cost-sensitive frequent polling. Shared security principle: deny scheduling tools on untrusted surfaces, self-contained prompts, minimal toolsets, and fail-loud on errors.

Agent Hermes & OpenClaw: Workspace Files and Prompt Assembly — A Deep Dive

最后更新 | Last updated: 2026-06-06

一、设计哲学对比 | Design Philosophy Comparison

中文

工作区文件是「文件即配置」理念的核心：Agent 的人格、流程与知识外化为 Markdown，在会话启动时组装进系统提示词。

English

Workspace files embody files-as-config: agent persona, procedures, and knowledge externalized as Markdown and assembled into the system prompt at session start.

二、OpenClaw 八文件 Bootstrap 体系 | OpenClaw Eight Bootstrap Files

中文

OpenClaw 在 agents.defaults.workspace（默认 ~/.openclaw/workspace/）中维护用户可编辑的 bootstrap 文件：

1
2
3
4
5
6
7
8
9
10
11
12
13

~/.openclaw/workspace/
├── SOUL.md           # 人格、语气、价值观、行为边界
├── AGENTS.md         # 操作手册：工作流、记忆规则、多 Agent 协作
├── USER.md           # 用户偏好与身份信息
├── TOOLS.md          # 工具使用指南（不控制工具是否存在）
├── IDENTITY.md       # Agent 名称、头像、元数据
├── HEARTBEAT.md      # 主动巡检任务清单
├── BOOTSTRAP.md      # 首次运行仪式（完成后删除）
├── MEMORY.md         # 长期记忆（Agent 可更新）
├── memory/           # 日记式记忆（按需，非自动注入）
│   └── 2026-06-05.md
└── skills/           # 技能目录（独立加载机制）
    └── weather/SKILL.md

2.1 注入顺序与优先级

新会话首轮流次，OpenClaw 将 bootstrap 文件注入 Project Context（系统提示词区块）：

设计意图：稳定的人格指令先于易变的记忆内容，降低记忆更新对行为核心的干扰。

2.2 体积控制

空白文件跳过；超大文件截断并附加标记，提示 Agent 用 read 获取全文。memory/*.md 不自动注入，通过 memory 工具按需读取。

2.3 BOOTSTRAP.md 特殊行为

• 仅当工作区无任何其他 bootstrap 文件时由 openclaw setup 创建
• 存在期间保持在 Project Context，指导首次仪式
• 完成后删除，不会在后续重启时重建
• 工作区 attestation 标记防止静默重种

English

OpenClaw maintains eight bootstrap files under the workspace. Injection order: SOUL → IDENTITY → USER → AGENTS → TOOLS → HEARTBEAT → BOOTSTRAP → MEMORY (most volatile last). Blank files skipped; large files truncated per bootstrapMaxChars (20k) and bootstrapTotalMaxChars (150k). memory/*.md is on-demand only. BOOTSTRAP.md is one-time for brand-new workspaces.

三、SOUL 与 AGENTS 职责分离 | SOUL vs AGENTS Separation

中文

两个框架均推荐将人格与流程拆分到不同文件：

反模式示例：

1
2
3
4
5
6
7

# ❌ SOUL.md 中写操作步骤
When deploying, run kubectl apply -f deploy.yaml...

# ✅ AGENTS.md 中写操作步骤
## Deploy workflow
1. Run tests first
2. kubectl apply with --server-side