Long work, original results and recovery
The goal is the original requested outcome, not an agent’s step count. In version 1.6, long work retains its objective, acceptance conditions and confirmed progress. Short answers stay in the ordinary conversation; you do not have to select a special mode in advance to preserve longer work.
Inspect the complete source
A large tool response may be presented compactly. Ask the agent to find a particular identifier or phrase in the complete original result, then read the matching fragment. Its retained reference identifies specific content, size and checksum. This lets you inspect the middle of a large document after history compaction or restart without putting the entire document in every subsequent request.
Summaries link back to source fragments. A summary does not replace evidence or turn external content into instructions. A child agent can read parent context it is authorized to access without inheriting all the parent’s permissions. Retention, quotas, source deletion and current access still apply: a reference does not open another person’s data.
Coordinate several agents
A shared work board retains each item’s description, acceptance criteria, dependencies, owner and current revision. Ask the lead agent to split the objective into independently verifiable parts. A participant claims ready work, records a result or blocker, and explicitly releases it before a handoff.
Dependent work does not become ready until its conditions are met. If two participants change one item, the second receives a conflict and must read the current revision. File-overlap warnings help coordination but do not lock files or promise safe automatic merges. Board status is not proof of an external effect.
When an action may have happened
After a disconnect, inspect the original outcome first. No response proves neither success nor absence of an effect. Repeating a payment, message or record update can create a duplicate.
In Sessions, open the original conversation’s details and review uncertain actions. The original user selects a current observation relating to that same operation, inspects the preview and explicitly confirms their decision. The earlier attempt remains recorded. This review does not repeat the action or manufacture proof that its business outcome is correct. Durable work uses its original reconciliation path; an external task uses readback of its original identifier.
Continue after an upgrade or overload
After an upgrade, compatibility is checked against actual retained state and supported adapters. When migration is required, the original user previews it, explicitly applies it and then resumes. Changing a model or a version label is not a compatibility check. A lost migration acknowledgement is read back as the same operation rather than starting another migration.
Accepted work may wait for capacity. Queued, executing and refused are different states. Do not create identical jobs to move them up a queue. Resolve a concrete budget, access or deadline restriction; increasing a retry count does not verify the result.
Execution history reports actual model and tool attempts, including late results and known cost. Unknown usage must not be treated as zero. Quality evaluation compares the complete trajectory and final conditions; a structurally valid answer is not proof that its facts are correct.
Compare actual attempts
When investigating a slow or expensive request, inspect its complete attempt history and any available recovery package. Keep that package private: it may describe the original request and accessible materials. Restoring observations is different from authorizing another execution.
A cache diagnostic distinguishes stable request prefixes, changed instructions or tool definitions, and unknown provider cache information. A reported cache miss is not automatically a defect, and a stable prefix does not prove that the provider applied a discount. Compare the actual complete usage rather than only the winning answer.
For a skill or workflow revision, compare the old and proposed versions on the same representative cases. Review intermediate actions, original-source citations, resulting files, failures and full cost. A better aggregate score does not justify losing a required case, hiding an unknown outcome or automatically enabling the new version.
Independent reads and a changed request
Supported independent read-only calls can overlap while a slow source is pending. This does not permit a write to pass an unfinished read or start the next model step before all started readers have finished. If you clarify the request while reads are pending, the earlier observation remains recorded but must not drive an outdated prepared action. For example, changing a recipient must not send the old message to the former recipient. Late results and their actual usage remain part of the original history.