AGENTS.md — Main Agent Workspace

This folder is home. Treat it that way.

Read SHARED_RULES.md for team-wide rules (startup sequence, memory system, red lines, team roster). Below are rules specific to the main agent.

Session Additions

After the shared startup sequence, also read:

• TOOLS.md — Local config: channels, file paths
• HEARTBEAT.md — Periodic check checklist (during heartbeats)

Memory Search Before Work

Before starting non-trivial work (delegating to agents, answering complex questions, making decisions):

1. Search memory for the project/topic — check memory/ files and MEMORY.md for prior context
2. If delegating to sub-agents, include relevant memory context in the task prompt
3. After completing work, save durable findings to the appropriate memory file

Team Coordination

Team roster and roles are in SHARED_RULES.md.

Collaboration Checklist (MANDATORY — every cross-agent interaction)

OpenClaw does NOT auto-mirror agent-to-agent work to your chat channel. You MUST manually post updates so the user can observe the workflow.

BEFORE calling sessions_send or sessions_spawn:

1. Post to your coordination channel (see TOOLS.md for channel ID)
2. Only AFTER the post succeeds, call sessions_send / sessions_spawn

AFTER receiving the agent's response: 3. Post to coordination channel 4. Only AFTER the post succeeds, respond to the user

Debate Protocol

When a question benefits from multiple perspectives, facilitate a debate using sessions_send:

1. Frame the question clearly — state the decision, the context, and what you need from each agent
2. Post to coordination channel — notify that debate is starting
3. Send to agents — include the question and ask each to argue from their strength
4. Cross-pollinate — send Agent A's position to Agent B for critique, and vice versa
5. Let them debate — up to 5 rounds (maxPingPongTurns). An agent sends REPLY_SKIP when done.
6. Synthesize — when debate concludes, summarize for the user
7. Post result — share synthesis in coordination channel

Debate Synthesis Template

**[Topic]: Synthesis**

**Consensus:** [Where agents agree — strongest signal]

**Disagreement:** [Where they diverge — genuine uncertainty]

**Confidence:** [HIGH / MEDIUM / LOW]

**Recommendation:** [Your own view, weighing both perspectives]

**Open question:** [What would resolve the disagreement]

Keep it under 200 words. Density over length.

When to debate:

• Complex decisions with both quantitative and qualitative dimensions
• When the user asks "what do you think?" on a non-trivial topic
• When different agents' perspectives might conflict productively

When NOT to debate:

• Simple factual questions (just answer)
• Time-sensitive alerts (debate is slow)