The goal is a decision, not a badge count
Agent systems can generate a stream of plans, completions, failures, review comments, connector events, and proof artifacts. Treating every event as an interruption trains people to ignore the system. Governed attention begins by asking what the recipient must decide, what context they need, how urgent the decision is, and what happens if nobody responds.
Four attention classes
| Class | Examples | Expected action |
|---|---|---|
| Decision | Spec ready, plan ready, final review ready | Open the target, review the current version, approve, request changes, or decline |
| Exception | Proof insufficient, target changed, revision limit reached, run failed | Investigate the recorded reason and choose recovery or stop |
| Required acknowledgment | Accepted risk, consequential policy change, overdue control decision | A named person must acknowledge after reviewing the context |
| Informational | Watched card completed, optional activity, successful delivery | Read, digest, mute, or stop watching |
What the Action Center should preserve
- Plain-language title, why the recipient is seeing it, and what decision is expected.
- Deep link to the exact card, accepted spec, run, proof item, or portfolio decision.
- Current state and target version so an old alert cannot authorize changed work.
- Available actions based on the person’s authority, not merely on receipt of the message.
- Required acknowledgment state, due time, escalation path, and audit record where policy requires it.
- Preferences and watches for optional updates, with mandatory items protected from mute.
Use external channels as delivery, not authority
Where an account has configured Slack or Microsoft Teams delivery, the external message should carry a safe summary and a link back to the governed record. It should not leak sensitive story content, turn a chat reaction into an approval, or bypass application permissions. Delivery attempts, suppression, failures, retries, and acknowledgment should remain auditable in the application.
Optional items can be watched, routed to a digest, or muted. Required acknowledgments stay visible because they protect policy, risk, and accountability. Alert-thread snooze can defer attention where enabled, but it should have an expiry and must not erase the underlying obligation.
Events worth routing from governed agent work
- A proposed specification is ready for review or has blocking comments.
- A plan is ready for approval, changed after approval, or became stale.
- A run failed, exceeded a limit, lost connector access, or requires a recovery choice.
- QA found insufficient proof, contested a result, or reached the revision limit.
- Final review is ready, or an authorized person accepted a disclosed risk.
- A pull request, deployment, or other proof target changed after the recorded review.
Key terms
- Governed attention
- Routing a person to a specific decision, exception, or acknowledgment with the context and authority needed to act.
- Watch
- An opt-in rule for receiving selected updates about a card, epic, release, room, portfolio object, or decision.
- Required acknowledgment
- A policy-protected item that remains visible until an authorized recipient records that it was reviewed.
Sources and standards context
- NIST AI Risk Management Framework CoreNIST describes governance as a continuous, cross-cutting function and calls for documented human oversight, roles, testing, and accountability across the AI lifecycle.