talonic_submit_agent_task

Return the result of a claimed Agent task. Output keys must match output_contract exactly, required fields must be present, and every value must match its declared data type. The platform validates the entire result before writing any cell.

Submit is the last call of the pull workflow — list, get, claim, heartbeat, submit — and the only one that changes the document. It requires the task_id plus the exact execution_epoch from your claim; a stale epoch is rejected with HTTP 409 just as it is on heartbeats, which is what makes the workflow safe to run from several workers at once.

{
  "task_id": "11111111-1111-4111-8111-111111111111",
  "execution_epoch": 3,
  "outputs": {
    "risk_score": {
      "value": 0.72,
      "confidence": 0.88,
      "reasoning": "Two adverse indicators in the supplied evidence."
    }
  },
  "summary": "Supplier requires enhanced review."
}

Validation is all-or-nothing. Every output key must exist in the task's output_contract, every required field must be present, and every value must match its declared data type. If any check fails, the platform rejects the whole submission and writes nothing — there is no partially-filled result to clean up. On success the task's status moves to submitted, the response returns the task's metadata, and the parked document resumes through the rest of its pipeline with your values in place.

The optional confidence (0–1) and reasoning (up to 4,000 characters) accompany each value into the document's audit trail, and summary describes the result as a whole. Filling them in is worth the tokens: reviewers see them next to the value when deciding whether to trust it, and they are the only trace of why the agent answered as it did once the conversation that produced the answer is gone.