How to run task agents
Submit typed inputs through the pre-filled run form or the API, and follow the run to a schema-validated output.
Task mode is a stateless invocation: typed input in, schema-validated output back, with no memory between runs and hard ceilings on duration and tool calls.
From the Console
- Open the agent in the Catalog and open the Run this agent panel.
- Edit the pre-filled JSON input — the skeleton matches the agent's declared input schema, so the required fields are already laid out.
- Submit and follow live progress on the run page.
- Collect the structured output when the run succeeds, or read the explicit error if it fails.
From your own systems
Products and workflows call the same API the Console uses: submit a run, stream or poll its progress, then fetch the result. Every UI action displays its curl equivalent, and Copy token in the user menu gives you a credential for testing. The UI is a view over the API — anything it does, your workflow or CI can do.
Bounded by design
Every run carries budgets: a maximum duration, a maximum number of tool calls, and a token budget. A run that exceeds them stops with an explicit error code rather than running open-ended.
Note
Task agents have no memory by design — every run starts clean. For multi-turn work, use a conversational agent (rolling out with the V1 Beta).
Still stuck?
Tell us what you were trying to do and where it failed — screenshots and run IDs help. A human follows up.