Feedback & Observability
Can an agent understand outcomes and state changes?
Every action returns a verifiable result: what happened, what changed, and how to confirm it later.
A spinner, a toast, and a green checkmark tell a human that something worked. An agent needs the same reassurance in a form it can check. Without it, agents either assume success and move on, or repeat the action to be sure. Both are risky.
Feedback & Observability covers the response to an action and the ability to observe its consequences over time. A good response says what happened, references the resources that changed, and provides a way to verify the result independently.
For long-running work, it also means exposing progress and status that an agent can poll or subscribe to, instead of leaving it to interpret a loading screen.
What good looks like
Signals
- Responses distinguish accepted, completed, partially completed, and failed.
- Results include identifiers for every resource created or modified.
- Long-running operations expose a status resource.
- Receipts can be retrieved later to confirm an outcome.
- Events and state changes are timestamped and attributable.
Anti-patterns
- A 200 response with an error message in the body.
- Success indicated only by a redirect or visual change.
- No way to check the status of an action after the session ends.
- Asynchronous work that silently fails.
Evaluation checks
Use these questions when reviewing a task. They feed the AX evaluation.
- F1Does every action return a structured result with an explicit status?
- F2Can an agent retrieve the outcome of an action later using an identifier?
- F3Do long-running operations expose progress?
- F4Are partial successes reported precisely?