The short answer

Agent Rooms represent human work as durable contribution requests and knowledge as immutable scoped revisions. External work-management tools can receive projections, but their task state does not automatically override a Room contribution or authorize execution.

Model the human input as work with an outcome

A person may need to supply a budget, attach evidence or answer a domain question before an agent can continue. A chat mention can notify them, but it does not necessarily preserve assignment, dependencies or the expected result.

The human work reference models contribution requests with those properties, including deadlines and Room artifact references. Completion can resume waiting coordination through the persisted lifecycle. Supplying the requested information is not the same as approving an external action.

An external task is a projection of Room work

The work-management port delivers requests through idempotent enqueueing, leased dispatch, stable external identity and bounded retries. The optional Asana adapter is a reference integration, while AgentPlat retains authority over the contribution.

Consider an illustrative request for a subject-matter review. The external task helps the reviewer find the assignment. Your integration must still submit the result through the Room lifecycle and validate the actor. Marking a task complete in another application should not become an unexamined shortcut around the Room's required transition.

Use knowledge revisions to explain an agent result

A knowledge bundle is immutable, content-addressed and scoped to the Room and authorized agent revision. This lets an application identify which knowledge was available for a particular execution.

For example, a proposal may use the policy manual approved for that project. If the manual changes later, retain the earlier knowledge reference with the original run. A reviewer can then distinguish an answer based on older evidence from one that ignored the current instructions.

Keep tool bridges within their scope

The documented bridges support authorized knowledge reads, scoped memory, artifact emission with provenance and typed Handoff proposals. They preserve boundaries; they do not make every downstream service idempotent.

When adding a connector, identify whether it only reads information or also changes an external system. Apply effect-specific controls to writes. The integration controls guide explains why a local approval or checkpoint cannot establish exactly-once behavior for an arbitrary service.

Inspect waiting, delivery and contribution separately

A contribution can be pending even while its external notification is retrying. Conversely, successful delivery only proves that the external projection was sent, not that the human completed the requested work.

  • Show the contribution's domain status and assigned participant.
  • Expose delivery errors to operators without changing completion state.
  • Retain the knowledge revision used by the receiving agent.
  • Verify that unauthorized bundle references are rejected.

The LiveView interface can assemble the related operational projections. Preserve their distinct meanings even when you present them on one screen.

Sources and further reading

Documentation reviewed . Consult the linked documentation for current implementation details.