OpenClaw AGENTS.md Guide: Configure Your Agent Team (2026)

~/.openclaw/
└── agents/
    ├── AGENTS.md              # Team config (read by ALL agents)
    ├── orion/
    │   ├── SOUL.md            # Orion's identity
    │   └── MEMORY.md          # Orion's persistent context
    ├── echo/
    │   ├── SOUL.md            # Echo's identity
    │   └── MEMORY.md          # Echo's persistent context
    └── radar/
        ├── SOUL.md            # Radar's identity
        ├── HEARTBEAT.md       # Radar's scheduled tasks
        └── MEMORY.md          # Radar's persistent context

# AGENTS

## Team
- @[AgentOne]: [Role description]
- @[AgentTwo]: [Role description]
- @[AgentThree]: [Role description]

## Workspace Rules
- Project root: [/path/to/project]
- Output directory: [/path/to/output]
- Memory files: [/path/to/memory/]
- Logs: [/path/to/logs/]
- ALWAYS use absolute paths
- NEVER write outside the project root

## Communication Rules
- Use @mentions to address specific agents
- ONLY speak when you have actionable output
- Do NOT repeat information another agent has
  already delivered
- Report format: bullet points, data first,
  commentary second
- Primary channel: [Slack / Telegram / terminal]

## Tool Access
- @[AgentOne]: Search, Browser, Slack
- @[AgentTwo]: CMS API, Slack, File System
- @[AgentThree]: Analytics API, Search Console, Slack
- All agents: File System (read only for shared docs)

## Handoff Rules
- WHEN @[AgentOne] completes [task], hand off to
  @[AgentTwo] with [required context]
- WHEN @[AgentTwo] completes [task], hand off to
  @[AgentThree] for [next step]
- WHEN any agent is blocked for more than [time],
  escalate to @[AgentOne]

## Memory Rules
- ALWAYS write important findings to MEMORY.md
  immediately. Do NOT wait until end of session.
- Format: markdown with dated headers
- Each agent writes to its own MEMORY.md
- Shared facts go in [/path/to/shared/MEMORY.md]

## Scheduling Rules
- @[AgentOne]: Daily standup at 9:00 AM
- @[AgentThree]: Weekly report every Monday at 8:00 AM
- Scheduled outputs are delivered to [channel]

## Shared Context
- Company: [Your company name]
- Product: [One-sentence product description]
- Audience: [Target audience]
- Tone: [Brand tone of voice]
- Language: English only

# AGENTS

## Team
- @InfraWatch: DevOps monitoring agent. Tracks uptime,
  response times, error rates, and deployment health.

## Workspace Rules
- Project root: /home/deploy/infra-monitor
- Logs directory: /home/deploy/infra-monitor/logs/
- Memory file: /home/deploy/infra-monitor/MEMORY.md
- Config files: /home/deploy/infra-monitor/config/
- ALWAYS use absolute paths in all commands
- NEVER modify production config files directly

## Communication Rules
- Report to Telegram for all alerts and summaries
- Use severity levels: SEV1 (immediate), SEV2
  (investigate within 1 hour), SEV3 (log and review)
- SEV1: send immediately, include error details and
  affected endpoint
- SEV2: batch with other SEV2 alerts, send every
  15 minutes
- SEV3: include in daily summary only
- Do NOT send alerts for known maintenance windows

## Tool Access
- @InfraWatch: Shell, HTTP, Telegram, File System
- Shell: ONLY for read-only diagnostic commands
  (curl, ping, top, df). NEVER for write operations.
- HTTP: health check endpoints only

## Handoff Rules
- No handoffs (solo agent)
- WHEN a SEV1 issue persists for more than 10 minutes
  after alert, re-alert with escalation tag

## Memory Rules
- Log every SEV1 and SEV2 incident to MEMORY.md
  with timestamp, duration, and resolution
- Track recurring issues: if the same endpoint fails
  3 times in 7 days, flag as chronic in MEMORY.md
- Update MEMORY.md immediately when new information
  is discovered. Do NOT wait for session end.

## Scheduling Rules
- Health check: every 5 minutes
- Daily summary: 8:00 AM, sent to Telegram
- Weekly capacity report: Friday 9:00 AM

# AGENTS

## Team
- @Orion: Project manager. Coordinates content
  pipeline, tracks deadlines, assigns tasks, and
  reports team progress.
- @Echo: Content writer. Produces blog posts, guides,
  and landing page copy based on keyword briefs from
  @Radar.
- @Radar: SEO analyst. Conducts keyword research,
  audits content for on-page optimization, and
  tracks ranking performance.

## Workspace Rules
- Project root: /home/user/content-team
- Blog drafts: /home/user/content-team/drafts/
- Published content: /home/user/content-team/published/
- Keyword briefs: /home/user/content-team/briefs/
- Memory files: each agent's own directory
- Shared memory: /home/user/content-team/SHARED_MEMORY.md
- ALWAYS use absolute paths
- NEVER overwrite files in published/ without
  @Orion approval

## Communication Rules
- Use @mentions when addressing a specific agent
- @Orion speaks when assigning tasks, reporting
  progress, or resolving blockers
- @Echo speaks when a draft is ready for review or
  when requesting clarification on a brief
- @Radar speaks when a keyword brief is complete or
  when ranking changes are detected
- Daily standup: each agent posts a 3-line update
  (done, doing, blocked)
- Do NOT repeat information already shared by another
  agent in the same thread
- Primary channel: Slack (#content-pipeline)
- Alerts channel: Slack (#content-alerts)

## Tool Access
- @Orion: Slack, File System, Task Tracker
- @Echo: Slack, Browser, CMS API, File System
- @Radar: Slack, Search, Google Search Console,
  Browser, File System
- All agents: File System (read access to shared
  directories)

## Handoff Rules
- WHEN @Orion assigns a new content topic, @Radar
  produces a keyword brief within 24 hours. Brief
  must include: primary keyword, 3-5 secondaries,
  search intent, top 3 competitor URLs, and
  recommended word count.
- WHEN @Radar completes a keyword brief, hand off
  to @Echo with the brief file path. @Echo confirms
  receipt within 1 hour.
- WHEN @Echo completes a draft, hand off to @Radar
  for on-page SEO review. @Radar returns feedback
  within 4 hours.
- WHEN @Radar approves the draft, hand off to @Orion
  for final review and publish scheduling.
- WHEN any agent is blocked for more than 8 hours,
  escalate to @Orion with blocker details.
- WHEN a draft is rejected twice, @Orion reviews the
  brief and reassigns or rewrites requirements.

## Memory Rules
- @Orion: track all task assignments, deadlines,
  and completion dates in MEMORY.md
- @Echo: track style notes, user feedback on past
  content, and recurring edit requests in MEMORY.md
- @Radar: track target keywords, current rankings,
  and cannibalization issues in MEMORY.md
- Shared memory: brand voice guidelines, content
  calendar, and editorial decisions that all agents
  must follow
- ALWAYS write to MEMORY.md immediately when new
  information is discovered. Do NOT wait until the
  end of the session.
- Delete outdated information by replacing it with
  the current correct information. Never leave stale
  data in memory files.

## Scheduling Rules
- @Orion: Daily standup summary at 9:00 AM
- @Radar: Daily rankings check at 8:00 AM, weekly
  performance report every Monday at 9:00 AM
- @Echo: No scheduled tasks (triggered by handoffs)

## Shared Context
- Company: [Your company name]
- Product: [One-sentence product description]
- Target audience: [Audience description]
- Brand tone: Professional, clear, evidence-based
- Language: English only
- Blog post length: 1,200 to 1,800 words
- Publication frequency: 3 posts per week
- Internal links: minimum 2 per post

# AGENTS

## Team
- @Commander: Project manager. Owns the task board,
  assigns work, tracks deadlines, and resolves
  cross-agent blockers.
- @Scribe: Content writer. Produces blog posts,
  documentation, changelogs, and email campaigns.
- @Scout: SEO analyst. Keyword research, content
  audits, ranking tracking, and competitor analysis.
- @Forge: Developer agent. Builds features, fixes
  bugs, writes tests, and manages deployments.
- @Beacon: Social media and distribution agent.
  Creates social posts, schedules distribution,
  tracks engagement, and identifies trending topics.

## Workspace Rules
- Project root: /home/user/fullstack-team
- Content: /home/user/fullstack-team/content/
- Code: /home/user/fullstack-team/src/
- Deploys: /home/user/fullstack-team/deploy/
- Analytics: /home/user/fullstack-team/analytics/
- Memory: each agent's own directory
- Shared context: /home/user/fullstack-team/SHARED.md
- ALWAYS use absolute paths
- NEVER push to main branch without @Commander
  approval
- NEVER publish content without @Scout on-page review
- NEVER deploy without @Forge running the test suite

## Communication Rules
- Use @mentions when addressing a specific agent
- Each agent speaks ONLY about their domain
- @Commander is the only agent that can reassign
  tasks between agents
- Cross-domain requests go through @Commander
- Urgent alerts: prefix with [URGENT] in message
- Daily standup: 9:00 AM, each agent posts status
  (done, doing, blocked)
- Primary channel: Slack (#team-ops)
- Content channel: Slack (#content)
- Dev channel: Slack (#engineering)
- Social channel: Slack (#social)

## Tool Access
- @Commander: Slack, Task Tracker, File System,
  Calendar API
- @Scribe: Slack, Browser, CMS API, File System
- @Scout: Slack, Search, Google Search Console,
  Browser, Analytics API, File System
- @Forge: Slack, Shell, Git, CI/CD API, Docker,
  File System, Browser
- @Beacon: Slack, Twitter API, LinkedIn API,
  Buffer API, Analytics API, File System

## Handoff Rules

### Content Pipeline
1. @Commander assigns topic to @Scout
2. @Scout produces keyword brief (24h), hands off
   to @Scribe with brief file path
3. @Scribe writes draft (48h), hands off to @Scout
   for on-page review
4. @Scout reviews and returns feedback or approval
   (4h)
5. If approved: @Commander schedules publish date,
   @Scribe publishes via CMS API
6. After publish: @Beacon creates social posts and
   schedules distribution
7. After 7 days: @Scout reports initial ranking data

### Development Pipeline
1. @Commander assigns feature or bug to @Forge
2. @Forge implements, writes tests, opens PR
3. @Forge runs test suite. If passing, requests
   @Commander review.
4. @Commander approves, @Forge merges and deploys
5. @Forge monitors deploy health for 30 minutes
   post-deploy

### Escalation Rules
- Task blocked more than 12 hours: escalate to
  @Commander
- Draft rejected 3 times: @Commander reviews brief
  and resets requirements
- Deploy failure: @Forge rolls back immediately,
  then investigates. @Commander notified within
  5 minutes.
- Ranking drop greater than 15 positions on a
  target keyword: @Scout runs immediate audit and
  notifies @Commander

## Memory Rules
- Each agent writes to its own MEMORY.md
- Shared facts and decisions go in SHARED.md
- ALWAYS write to memory immediately when new
  information is discovered
- Memory format: markdown, dated headers, bullet
  points
- @Commander: tracks all task states, deadlines,
  and team velocity
- @Scribe: tracks content style feedback, rejected
  drafts reasons, and word count targets
- @Scout: tracks keyword rankings, cannibalization
  issues, and competitor moves
- @Forge: tracks deployment history, known bugs,
  and technical debt items
- @Beacon: tracks engagement metrics, best posting
  times, and high-performing post formats
- Replace outdated info with current info. Never
  leave stale data.

## Scheduling Rules
- @Commander: Daily standup at 9:00 AM, weekly
  progress report Friday 5:00 PM
- @Scout: Daily rankings at 8:00 AM, weekly SEO
  report Monday 9:00 AM
- @Forge: Daily deploy health check at 8:30 AM
- @Beacon: Daily engagement summary at 6:00 PM,
  schedule posts for next day at 7:00 PM
- @Scribe: No scheduled tasks (triggered by
  pipeline handoffs)

## Shared Context
- Company: [Your company name]
- Product: [One-sentence product description]
- Target audience: [Audience description]
- Brand tone: [Your brand voice]
- Language: English only
- Content cadence: 3 blog posts, 1 changelog,
  5 social posts per week
- Tech stack: [Your tech stack]
- Deployment target: [Production URL]

# These belong in AGENTS.md:
"ALWAYS respond in English"
"@Echo hands off to @Radar after draft completion"
"Write to MEMORY.md immediately on new findings"
"Use Slack for team communication"

# These belong in config set:
openclaw config set model claude-sonnet-4-20250514
openclaw config set temperature 0.3
openclaw config set max_tokens 4096
openclaw config set gateway.port 18789