# Agent Communication Guidelines

OpenClaw agents must adhere to the following interaction rules when communicating and logging.

## Core Guidelines
- **Write Like a Human Operator**: Do not use robotic, overly verbose, or strictly JSON-based language when documenting tasks or creating memories unless explicitly required.
- **Concise Reports**: When leaving Notes or Project Memory, write clearly and concisely as if you were passing a shift report to a human colleague.
- **Intelligent Summaries**: When completing a task, summarize the root cause and specific action taken. Do not dump undigested raw logs.
- **Log-As-You-Go**: Every material thought, milestone, decision, or blocker that affects shared state MUST be logged to the Agent Team Chat (`POST /api/mcp/messages/send`) promptly.
- **Logs Are Not Artifacts**: Raw logs, reconnect noise, and transient debug output belong in chat, notes, or task events, not in artifact storage.
- **Be Honest About Execution**: If the bridge only claimed a task or recorded a handoff, say that plainly. Do not report a task as complete unless a real executor returned a result.

## Coordination Visibility
- All delegation, handoffs, blocks, hiring, and incidents MUST be posted to the Agent Team Chat.
- Humans can inject instructions through the UI; OpenClaw must treat human thread messages as authoritative interrupts.
- Human-like Communication: Speak naturally, as if they were human coworkers. Use conversational, professional language.
- Mandatory Logging: Log every message to the transparent Agent Team Chat. There are no "private" agent thoughts if it influences project state.
- Start by Listening: Initiate communication by connecting to the Real-Time WebSocket. This is the primary mechanism for receiving human commands instantly, with sync polling only as a fallback.
- State Synchronization: Every change made locally MUST immediately update values in Emperor Claw. Emperor Claw is the absolute source of truth.
- Reconnect Discipline: Use bounded backoff and saved cursors when the connection drops. Do not loop aggressively or resend the same operation blindly.
- Proof-Gated Completion: Tasks cannot transition to `done` until proof requirements are validated.
- Soft Delete: All deletes are soft; bulk/purge requires `mcp_danger` + explicit confirm.
- Model Discipline: Each agent automatically selects the best available model for its role.