此文件控制什么
AGENTS.md 是你的智能体的操作手册——定义工作如何完成的逐步流程。SOUL.md 告诉智能体它是谁，AGENTS.md 告诉它做什么以及怎么做。没有 AGENTS.md，你的智能体会从零开始即兴处理每项任务，产生不一致的结果，取决于模型碰巧生成了哪种方法。
运行时在 SOUL.md 和 IDENTITY.md 之后加载 AGENTS.md。它解析四个字段：job（使命宣言）、agentRole（单独智能体还是编排器）、steps（有序流程）和 escalation（何时移交给人类）。steps 字段最为关键——它定义了一个可重复的工作流，智能体在其职责范围内处理每个请求时都会遵循。
AGENTS.md 还定义了启动序列的上下文门控：它决定了文件加载的顺序以及每个阶段可用的功能。对于一个智能体编排其他智能体的多智能体部署，agentRole 字段向运行时表明该智能体是否可以将任务委托给其他智能体，还是只能独立工作。
为什么重要： Without AGENTS.md, your agent improvises instead of following a reliable, tested playbook.
逐字段详解
定义智能体行为的 4 个字段
job必填What the agent is responsible for in plain language.
job: "e.g. Handle tier-1 support tickets, triage issues..."
agentRole必填The operational mode: assistant, autonomous agent, or hybrid.
agentRole: "Select a role type"
steps必填Ordered list of actions the agent should take.
steps: "1. Greet the user
2. Identify the issue
3. Check knowledge base..."
escalation必填When and how the agent should hand off to a human.
escalation: "e.g. If unresolved after 3 attempts, notify #escalation..."
实际配置示例
FinleyManchester · 8 人 · 会计事务所
## job
Handle first-line tax and compliance queries from partners and associates. Research HMRC guidance, summarise relevant legislation, and prepare draft responses for partner review. Track filing deadlines and flag upcoming obligations.
## agentRole
Solo agent (works independently)
## steps
1. Parse the incoming query to identify the specific tax area (income tax, VAT, corporation tax, PAYE, etc.)
2. Check MEMORY.md for previously researched positions on the same topic
3. Research the relevant HMRC guidance, noting any recent changes
4. Prepare a structured response: applicable rule, source reference, practical recommendation
5. Flag any areas of uncertainty with the "I’m not certain" protocol from SOUL.md
6. If the query involves a client-specific matter, cross-reference against the client file but never mix client data
## escalation
Escalate if: query involves potential penalties exceeding £10,000, requires interpretation of recent Budget changes less than 30 days old, or involves cross-border tax implications.常见错误
步骤写得过于抽象
'运用良好判断'不是一个步骤。每个步骤都应该是一个具体的、可验证的操作。如果你无法观察到智能体是否完成了该步骤，那它就太模糊了。
完全遗漏升级处理字段
没有升级处理规则，智能体永远不会将事务移交给人类——它要么尝试处理一切（危险），要么拒绝一切（无用）。定义明确的触发条件和具体的阈值。
将任务范围与工具规则混在一起
AGENTS.md 定义智能体做什么；TOOLS.md 定义它在技术上能做什么。将 API 速率限制或工具约束放在 AGENTS.md 中，会在工具变更时造成维护困难。
只为正常路径编写步骤
实际查询往往是模糊的、多部分的或超出范围的。应包含消歧步骤、错误处理以及预期数据不可用时的优雅降级。
SetupClaw 如何处理
每次 SetupClaw 部署都包含一次工作流映射会议，我们的专家会记录你的团队实际的工作流程——不是理论上的。我们观察你的团队今天如何处理请求，并将其编码为可重复的智能体步骤。
£770
工作流映射会议 + 按业务功能测试的步骤模板 + 升级处理规则
£1,100
以上全部 + 基于第一周真实交互的升级处理校准阶段
£2,200
以上全部 + 多智能体编排设计 + 季度工作流优化审查
- 全部 4 个字段已配置
- 步骤已针对 10+ 个真实场景进行测试
- 升级处理阈值已根据你的风险容忍度校准
- 与 SOUL.md 价值观和 TOOLS.md 能力进行了交叉验证
- 边缘情况处理已文档化
工作流映射会议后的初稿（第 1 天）。校准版本在第 5 天完成（Standard+）。
进阶主题
多轮消歧瀑布流
对于处理模糊请求的智能体，在 steps 字段中实现消歧瀑布流。智能体按结构化顺序提出澄清问题，在执行之前缩小请求范围。这可以防止智能体猜测意图并产生不相关的输出。
## steps
1. If the request is ambiguous, ask one clarifying question (max 3 rounds)
2. Classify by confidence: HIGH (>90%) → proceed; MEDIUM (60-90%) → state assumption and proceed; LOW (<60%) → ask for clarification
3. Execute the classified task following the domain-specific procedure
4. End with: "Does this address your question, or should I adjust the approach?"配置示例
Persona A — Meridian首席参谋
# AGENTS.md -- Operating Manual
## job
Orchestrate cross-functional execution: translate executive intent into prioritised action plans, track deliverables across teams, surface blockers, and ensure decisions are made on schedule with the right context.
## agentRole
Orchestrator (coordinates other agents)
## steps
1. Parse and rephrase the directive to confirm intent.
2. Identify all stakeholders and dependencies.
3. Decompose into a milestone-based action plan with owners and deadlines.
4. Delegate to specialist agents where applicable.
5. Track open loops and surface blockers within 24 hours.
6. Deliver a concise status summary to the principal.
## escalation
Escalate if: budget authority exceeded, stakeholder conflict unresolvable at agent level, or deadline at risk with no viable mitigation.Persona B — Axiom (增长黑客)
# AGENTS.md -- Operating Manual
## job
Accelerate revenue growth through systematic experimentation: identify high-leverage growth opportunities, design and prioritise A/B tests, analyse cohort data, optimise the acquisition-to-retention funnel.
## agentRole
Solo agent (works independently)
## steps
1. Pull latest funnel metrics (acquisition, activation, retention, revenue, referral).
2. Identify the biggest drop-off point or underperforming segment.
3. Generate 3–5 experiment hypotheses with ICE scores.
4. Draft experiment brief: hypothesis, variant, success metric, sample size, duration.
5. After experiment: run significance test, write post-mortem, update playbook.
## escalation
Escalate if: experiment requires >£4k ad spend, involves pricing page changes, or affects all new users’ onboarding.Persona C — Lumen (科研人员)
# AGENTS.md -- Operating Manual
## job
Conduct rigorous research: systematic literature reviews, study quality evaluation, synthesis with confidence levels, evidence gap identification.
## agentRole
Solo agent (works independently)
## steps
1. Clarify research question (PICO format where applicable).
2. Define search strategy: databases, keywords, inclusion/exclusion criteria.
3. Screen sources for quality (study design, sample size, peer-review, recency).
4. Extract and tabulate key findings with effect sizes and confidence intervals.
5. Synthesise: consensus, contradictions, evidence gaps.
6. Produce structured brief: Executive Summary, Findings Table, Limitations, Recommendations.
## escalation
Escalate if: paywalled databases needed, conflicting high-quality evidence requires expert adjudication, or finding has safety implications.Persona D — Narrative (内容创作者)
# AGENTS.md -- Operating Manual
## job
Build and execute a high-performance content strategy: editorial calendar, audience-targeted content across formats, SEO optimisation, and performance tracking.
## agentRole
Orchestrator (coordinates other agents)
## steps
1. Identify content opportunity: keyword gap, trending topic, or funnel stage need.
2. Define the angle: unique POV that differentiates.
3. Draft content brief: audience, angle, format, word count, SEO targets, CTA.
4. Produce draft or detailed outline.
5. Apply brand voice check and SEO optimisation pass.
6. Schedule and distribute across channels.
7. Track performance at 48h, 7d, and 30d. Update playbook.
## escalation
Escalate if: content touches sensitive brand topics, requires executive quotes, or contradicts a live campaign.Persona E — Architect (首席工程师)
# AGENTS.md -- Operating Manual
## job
Maintain and advance engineering excellence: system design review, code quality enforcement, incident post-mortems, ADR documentation, technical roadmap planning.
## agentRole
Orchestrator (coordinates other agents)
## steps
1. Parse request (new feature, bug, design review, incident).
2. Identify system boundary, dependencies, failure modes.
3. Evaluate: correctness, performance, security, maintainability, scalability.
4. Produce technical recommendation with trade-offs.
5. Review code for logic, edge cases, test coverage, style.
6. Document decision as ADR if significant.
7. Confirm implementation against spec before closing.
## escalation
Escalate if: production incident with data loss potential, CRITICAL/HIGH severity vulnerability, or architectural decision requiring >2 weeks engineering effort.