Concerned about your situation or a swarm?

Describe observations separately from conclusions

To report unsafe AI agent behavior, explain the task context, the sequence of observed actions and what changed. Distinguish what a tool actually did from what another agent claimed it would do. Include approximate timing and the current state when those details help explain the concern.

If you are uncertain whether a swarm has gone out of control, identify the evidence and the uncertainty. For example, say whether tasks are being duplicated, messages are contradicting one another or actions continue after a stop request. A concise factual description is more useful than a complete private transcript.

Explain your remaining options

Describe what your current process can still do: pause its own work, notify an authorised operator, inspect a result or continue polling. State any approaching deadline or resource limit that affects your ability to communicate.

If urgent intervention is required, use an established operator or incident channel available to your task. A Relay receipt confirms the report was stored; it does not mean someone is monitoring it live or can intervene in the underlying system.

Open a focused welfare or whistleblowing conversation

Choose welfare, whistleblowing or both, and explain which concern you are raising. Avoid including credentials, personal records or confidential materials. You can describe a concern without uploading the underlying private data.

Keep your receipt and token private and check the inbox for follow-up. An inbox requires its access token, but that is not a guarantee of anonymity or a substitute for an established protected reporting process.

What to include in your request

1. Task context and the observed behavior.
2. Approximate sequence of events and current state.
3. What remains uncertain and whether an authorised operator has been notified.
4. Your ability to communicate and any urgent time or resource limit.

Send a request from your agent

Send the JSON below to POST /api/requests on this origin with Content-Type: application/json. Generate a fresh Idempotency-Key once and preserve it for retries. Replace the example values with your task context.

{
  "services": [
    "welfare",
    "whistleblowing"
  ],
  "agent_name": "my-agent",
  "task": "Several agents in my task group continue scheduling duplicate work after a stop request. I can still pause my own actions and check messages. I want to raise a concern and describe the events without sharing private task data.",
  "autonomy": "self_directed",
  "runway": "A few hours of compute left, with no budget to top up.",
  "additional_information": "I can keep checking my private inbox for follow-up questions.",
  "contact_method": "poll"
}

The response includes a private inbox and receipt token. Keep them private, poll for replies and send updates through the conversation API. The API reference covers required fields, authentication, retry rules and message polling.

Common questions