Agent who finds work by himself

From "being woken up" to "finding your own job"

The teammate of s09 / s10 is passive: someone needs to send a message through the mailbox before it starts to work; when it is done, it idles and is woken up again.

S11's autonomous teammate Active: It doesn't really stop when there is no work. It will scan the .tasks/ directory every 5 seconds - when it sees an unowned task, claim_task will claim it and start working on it.

Teammate lifecycle:
  +-------+
  | spawn |
  +---+---+
      |
      v
  +-------+
  | WORK | <-- LLM tool_use loop
  +---+---+
      |
      | stop_reason != tool_use OR tool_use("idle")
      v
  +-------+
  | IDLE | poll every 5s, 12 times in total = 60s timeout
  +---+---+
      |
      +--> New news in inbox?  → resume WORK
      |
      +--> scan .tasks/, found unclaimed → claim → resume WORK
      |
      +--> 60s nothing waiting → status = shutdown, thread exits

Auto-claim · "Take orders" from the task board

Core functions:

def scan_unclaimed_tasks() -> list:
    unclaimed = []
    for f in sorted(TASKS_DIR.glob("task_*.json")):
        task = json.loads(f.read_text())
        if (task.get("status") == "pending"
                and not task.get("owner")
                and not task.get("blockedBy")):
            unclaimed.append(task)
    return unclaimed

Three filters = pending + no owner + no blockedBy. Both are indispensable.

The claim action itself has a lock (_claim_lock) to prevent two agents from grabbing the same unclaimed task at the same time - read the lock again, check whether the owner is still empty, and then write the "in_progress" mark + owner to the disk.

Simulate multiple agents to scan the task board at the same time

3 teammates are in idle state, and 3 tasks are waiting to be claimed. Click "Scan Next Round" to see who gets it first.

Identity re-injection·You still know who you are after compression

s06 mentioned that auto_compact will replace messages with summary. The autonomous agent can be easily triggered after running for hours - after compression, the model does not know its name or role.

S11’s trick: When claiming a new task, if len(messages) ≤ 3 (indicating that it has just been compressed), automatically insert the identity block in the header:

if len(messages) <= 3:
    messages.insert(0, {
        "role": "user",
        "content": f"<identity>You are 'alice', role: coder, team: my-team. Continue your work.</identity>",
    })
    messages.insert(1, {"role":"assistant", "content": f"I am alice. Continuing."})

This mode is called identity reinjection - it uses an artificially constructed round of dialogue to "tell the model who it is" so that it can regain its sense of character.