OpenClaw ACP Harness Explained — How to Run Coding Agents in OpenClaw

What Is the ACP Harness?

ACP stands for Agent Control Protocol — OpenClaw's system for spawning, managing, and communicating with specialized sub-agents. The ACP harness is what lets your main OpenClaw agent spin up a Claude Code or Codex session, give it a coding task, and receive the results — all programmatically, without you doing anything manually.

Think of it as your agent hiring a specialized contractor for specific work, then reviewing and merging the results.

How the ACP Harness Works

1. Main agent identifies a coding task — something too complex for inline handling
2. Main agent calls sessions_spawn with runtime="acp" and the coding agent ID
3. ACP harness launches the coding agent — Claude Code, Codex, or another configured agent
4. Coding agent completes the task — writes files, runs tests, creates commits
5. Results auto-announce back to the main agent — no polling needed
6. Main agent synthesizes and reports — summarizes to you in Slack

Spawning a Claude Code Session

# How your main agent spawns a Claude Code sub-agent:
sessions_spawn({
  runtime: "acp",
  agentId: "claudecode",
  task: """
    Fix the bug in src/payments/stripe.ts:
    - Function processWebhook() throws on null customer ID
    - Add null check before line 47
    - Run existing tests to verify fix
    - Commit: fix: handle null customer in webhook processor
  """,
  model: "anthropic/claude-sonnet-4-6",
  timeoutSeconds: 1800
})

The Task Brief Matters

The most important part of ACP harness usage is the task brief. The coding agent shouldn't need to explore — it should get a blueprint:

## Good ACP Task Brief
- Exact files to modify (with line numbers if possible)
- What to add, change, or remove — be surgical
- What to preserve (regression risks)
- Tests to write or run
- Commit message format
- Deploy command if applicable

## Bad ACP Task Brief
"Fix the login bug" (too vague)
"Improve the dashboard" (no specifics)

Thread-Bound ACP Sessions

For ongoing work in a Discord or Slack thread, spawn ACP sessions with thread:true:

sessions_spawn({
  runtime: "acp",
  agentId: "claudecode",
  thread: true,
  mode: "session",
  task: "Help refactor the authentication module"
})

The coding agent maintains conversation history in the thread — you can guide it interactively.

Multiple Parallel ACP Sessions

Your main agent can spawn multiple ACP sessions in parallel:

## Parallel Development Pattern (AGENTS.md)
For large features with independent components:
1. Spawn ACP session 1: implement API endpoints
2. Spawn ACP session 2: implement frontend components  
3. Spawn ACP session 3: write integration tests
4. Wait for all three to complete
5. Review results, resolve any conflicts, create PR

Model Selection for ACP Sessions

Always specify the model in sessions_spawn:

model: "anthropic/claude-sonnet-4-6"  # ← Use Sonnet for sub-agents
# Never omit model — defaults may use expensive Opus unnecessarily

Ready to put this into practice? The OpenClaw Playbook has step-by-step walkthroughs, copy-paste configs, and real-world automation recipes. Get it for $9.99 and build your AI-powered setup today.