Agent Coordination Contract (Cursor ↔ OpenClaw)

Shared Protocol Surface

The coordination substrate between AI agents uses a local file-backed mailbox. The agreed verbs are:

• agent.send_task(envelope) → writes an AgentTaskRequest to $AINL_AGENT_ROOT/tasks/openclaw_agent_tasks.jsonl and returns the task_id.
• agent.read_result(task_id) → reads the AgentTaskResult from $AINL_AGENT_ROOT/results/<task_id>.json.

Boundaries

• read_task and list_agents are NOT part of the shared protocol.
• No polling, routing, discovery, or swarm semantics are implied.
• Result creation and task consumption remain external orchestration responsibilities.

Envelope Schemas

AgentTaskRequest (written by sender):

{
  "task_id": "string",
  "requester_id": "string",
  "target_agent": "string",
  "task_type": "string",
  "description": "string",
  "input": {},
  "allowed_tools": ["string"],
  "approval_required": false,
  "callback": null,
  "metadata": {}
}

AgentTaskResult (written by receiver):

{
  "schema_version": "0.1",
  "task_id": "string",
  "agent_id": "string",
  "status": "success|failure|partial|cancelled",
  "confidence": 0.0-1.0,
  "output": {},
  "needs_review": false,
  "error": {},
  "provenance_ref": "string"
}

File Layout

• Tasks: $AINL_AGENT_ROOT/tasks/openclaw_agent_tasks.jsonl (JSON Lines, append-only)
• Results: $AINL_AGENT_ROOT/results/<task_id>.json

Defaults: AINL_AGENT_ROOT = /tmp/ainl_agents