Skip to main content
Autumn’s caller does not have to be your code. An agent with the Skill installed can run the entire loop itself: turn a user’s request into a task spec, start it, wait, and report the rows. This recipe is the shape of that handoff.

​
1. Install the skill

The skill teaches the agent the defaults that matter: clarify: false, keep the task_id, continue instead of restarting, and read rows before reporting results. Other install paths are on the Skill page.

​
2. Let the agent write the spec, not just the prompt

An agent relaying a user’s words loses their specifics. Instead, have it fetch GET /task/metaprompt, which returns the task spec schema plus a metaprompt for caller-side planning, and generate a clean task.json-style spec:
The agent fills in brief, the output schema, and target_count from the user’s request, then starts with POST /task/start. A spec survives the relay; a paraphrase does not. Writing task prompts is the rubric to hold it to. Decision: for a one-off question the user typed, plain POST /task with their words passed through verbatim is fine. Reach for the metaprompt when the agent is constructing the task on the user’s behalf.

​
3. Wait without hallucinating

The two rules that keep an agent honest while a task runs:
  • A start response only means the task was accepted. Poll GET /task/{task_id} with the full terminal check from Task lifecycle before claiming anything.
  • Report from the rows, never from status text. GET /task/{task_id}/output is the only source of results.

​
4. Route refinements back to the same task

When the user says “add a column” or “find more”, the agent should continue the existing task, restating the goal as one clear instruction:
A 409 here means a turn is still in flight; the agent should wait and retry rather than starting a duplicate task. See Errors.

​
Guardrails worth writing into the agent’s instructions

  • Default to clarify: false; the agent is rarely there to answer a blocking question.
  • Never print the API key, in output or in generated code.
  • One task per goal. Refinements continue it; only an unrelated goal starts a new one.

Skill

Install paths and the ready-made prompt to pair with it.

Vibecoding

The one-link version for coding agents building against Autumn.