talonic_get_agent_task
Fetch the immutable input snapshot captured when one document reached an Agent stage. The payload also carries stage instructions, timeout policy, and the only output fields the agent may return.
The payload extends the metadata from talonic_list_agent_tasks with four fields. input_snapshot is the frozen view of the document's extracted data at the moment it entered the stage — later edits to the document do not leak into a running task. output_contract declares every field the agent may submit, each with a key, a dataType, and a required flag. instructions is the stage author's free-text brief. timeout_fallthrough (hold, skip, or route_to_review) says what happens to the document if nobody submits before timeout_at.
Fetching a task does not lease it: another agent can still claim the work while you are reading it. Treat get as the inspection step — confirm the instructions and contract are something you can actually fulfil — then take the lease with talonic_claim_agent_task before doing any real processing. If a claim races you to it, pick another task from the list rather than waiting on that one.
| Parameter | Type | Description |
|---|---|---|
| task_id * | UUID | Task ID returned by `talonic_list_agent_tasks`. |
{ "task_id": "11111111-1111-4111-8111-111111111111" }Payload access is deliberately narrow. The worklist never includes snapshots, and each talonic_get_agent_task call is recorded as an audited disclosure of that document's data to the calling credential. Fetch payloads for tasks you actually intend to process rather than sweeping the queue — it keeps the audit trail meaningful and avoids disclosing document contents your workflow never needed.
{
"id": "11111111-1111-4111-8111-111111111111",
"status": "available",
"input_snapshot": { "vendor_name": { "value": "Acme" } },
"output_contract": [
{ "key": "risk_score", "dataType": "number", "required": true }
],
"instructions": "Assess supplier risk from the supplied evidence.",
"timeout_fallthrough": "route_to_review"
}