Guides

Guides are end-to-end walkthroughs for the integration shapes we see most. Each assumes you have a key and have made one successful run.

Attaching context

Context sources are opt-in per run. Pass context: ["documents"] and the runtime resolves relevant items through the graph — by meaning, not keyword — fetching each through the kernel with the scope your key holds. The trace records every item fetched; nothing is read silently.

{
  "model": "y0-fast",
  "prompt": "Draft a renewal reminder for the Meridian contract.",
  "context": ["documents", "calendar"],
  "max_steps": 4
}

Constraining outputs

Pass output_schema and the run is validated against it internally — malformed output is retried inside the run rather than returned. Schema-exact extraction is benchmarked at /models/benchmarks.

Going from runs to agents

When a run needs to act — send, file, book — switch to the agents family. Declare tools up front, set approval gates on consequential actions, and keep max_steps tight. The kernel blocks any action the key's scopes do not grant, no matter what the model plans.

Putting evaluation in the loop

Before shipping a prompt change, run your eval suite (y0-judge against a frozen rubric) over a pinned case set. In production, sample runs continuously and alert on drift. The evaluation family docs cover rubric packs and calibration.