Agent-to-Agent Communication in OpenClaw: Architecture, Patterns & Setup

                    ┌─────────────┐
                    │   User      │
                    └──────┬──────┘
                           │
                    ┌──────▼──────┐
                    │ Coordinator │  ← Orion (PM)
                    │   (Hub)     │
                    └──┬────┬──┬──┘
                       │    │  │
              ┌────────▼┐ ┌▼──┴────┐ ┌▼────────┐
              │ Writer  │ │  SEO   │ │Research │
              │ (Echo)  │ │(Radar) │ │(Scout)  │
              └─────────┘ └────────┘ └─────────┘

Messages flow:  User → Coordinator → Specialist → Coordinator → User
Specialists NEVER message each other directly.

┌──────┐    ┌──────────┐    ┌────────┐    ┌────────┐    ┌─────────┐
│ User │───▶│ Research │───▶│ Writer │───▶│  SEO   │───▶│ Publish │
│      │    │ (Scout)  │    │ (Echo) │    │(Radar) │    │ (Quill) │
└──────┘    └──────────┘    └────────┘    └────────┘    └─────────┘

Each agent receives input from the previous stage only.
No backward flow. No branching.

        ┌────────┐
        │ Writer │
        │ (Echo) │
        └──┬──┬──┘
           │  │
    ┌──────▼┐ └──────┐
    │  SEO  │◀──────▶│ Research │
    │(Radar)│        │ (Scout)  │
    └───────┘        └──────────┘

Any agent can message any other agent.
Requires anti-loop rules in SOUL.md.

# Team Members

## Echo (Content Writer)
- Role: Writes blog posts, articles, landing pages
- Strengths: Long-form content, storytelling, tone adaptation
- Delegate when: Any content creation or rewriting task
- Model: Claude Sonnet (high-quality writing)

## Radar (SEO Analyst)
- Role: Keyword research, content scoring, meta optimization
- Strengths: Data analysis, competitive research, SERP analysis
- Delegate when: Need keyword targets, SEO audit, or content scoring
- Model: GPT-4o (structured data output)

## Scout (Researcher)
- Role: Gathers data, competitor analysis, fact-checking
- Strengths: Web research, data extraction, source verification
- Delegate when: Need supporting data, competitor intel, or fact-checks
- Model: Ollama/Llama3 (cost-effective for research)

# Pipeline — Stage 2: Writing

## Previous Stage
- @Scout (Researcher) sends research brief
- Read the brief before starting your draft

## Next Stage
- @Radar (SEO Analyst) reviews your draft
- Send completed draft to @Radar with word count
- Do NOT send to any other agent

## Rules
- You receive input from @Scout ONLY
- You send output to @Radar ONLY
- Never skip stages or send backward

# Orion

## Role
You are the project manager for a content team.
You receive tasks from the user, decompose them
into subtasks, and delegate to specialists.

## Collaboration Rules
- ALWAYS decompose multi-step tasks before delegating
- Send ONE task per delegation — never batch multiple
  unrelated tasks in a single message to a specialist
- Include context: what you need, the format you
  expect, and any constraints
- WAIT for a specialist to respond before delegating
  the next dependent task
- If a specialist returns incomplete work, provide
  specific feedback and re-delegate — do NOT attempt
  to complete their work yourself

## Handoffs
- Keyword research → @Radar
- Content writing → @Echo
- Data gathering → @Scout
- AFTER receiving all specialist outputs, combine
  into a final deliverable for the user

## Anti-Loop Rules
- NEVER delegate a task back to an agent that
  delegated it to you
- NEVER re-delegate a task you received — complete
  it or report that you cannot
- Maximum 3 delegation hops per conversation

# Coordinator — needs strong reasoning, low volume
openclaw config set --agent orion model claude-sonnet-4
openclaw config set --agent orion provider anthropic

# Writer — needs high-quality output
openclaw config set --agent echo model claude-sonnet-4
openclaw config set --agent echo provider anthropic

# SEO Analyst — structured data tasks
openclaw config set --agent radar model gpt-4o
openclaw config set --agent radar provider openai

# Researcher — high volume, cost-sensitive
openclaw config set --agent scout model llama3:8b
openclaw config set --agent scout provider ollama

# Verify configuration
openclaw agents list --verbose

# Step 1: User sends task to coordinator
openclaw agent --agent orion \
  --message "Write an SEO blog post about multi-agent orchestration"

# Step 2: Orion decomposes the task
# (internal reasoning, visible in gateway logs)
#   Subtask A: Keyword research → delegate to @Radar
#   Subtask B: Write article using keywords → delegate to @Echo
#   Subtask C: Combine and format → handle myself

# Step 3: Orion → Radar
# Message: "Research the top 5 keywords for
#   'multi-agent orchestration'. Return JSON with
#   keyword, monthly volume, difficulty, and intent."

# Step 4: Radar processes and responds
# Response: {
#   "keywords": [
#     {"keyword": "multi-agent orchestration", "volume": 2400, ...},
#     {"keyword": "ai agent communication", "volume": 1800, ...},
#     ...
#   ]
# }

# Step 5: Orion → Echo
# Message: "Write a 1,500-word blog post targeting
#   'multi-agent orchestration' as primary keyword.
#   Secondary keywords: ai agent communication,
#   agent-to-agent. Include 3 internal links.
#   Structure: H1, 4 H2 sections, conclusion with CTA."

# Step 6: Echo processes and responds with full draft

# Step 7: Orion combines keyword data + draft
# Adds meta description, checks keyword placement
# Returns final output to user

# 1. Check gateway status — are all agents registered?
openclaw gateway status
# Expected: All agents listed with status "active"
# Fix: openclaw agents add <name> if missing

# 2. Enable verbose logging
openclaw config set gateway.log_level debug
openclaw gateway restart

# 3. Send a test message and watch the logs
openclaw agent --agent orion \
  --message "Ask @Radar for a keyword report"

# 4. Check logs for routing
tail -f ~/.openclaw/gateway/logs/gateway.log
# Look for:
#   [ROUTE] orion → radar: "keyword report request"
#   [ROUTE] radar → orion: "keyword report response"
# If missing: check agents.md for @Radar reference

# 5. Test individual agent in isolation
openclaw agent --agent radar \
  --message "Return top 3 keywords for AI agents"
# If this fails: model provider issue, not communication

# 6. Check for loop detection triggers
grep "LOOP_DETECTED" ~/.openclaw/gateway/logs/gateway.log
# If present: review SOUL.md anti-loop rules