Plans and acceptance

A plan shows intended work and progress. Acceptance criteria check the result. Milestones report declared events during the work. None of these replaces the execution policy or a workflow approval.

Make progress visible with a plan

In Input & output → Run behavior, enable Keep a plan to let the agent list steps and mark its progress. This requires a model with tool support. The run's Plan view lets people follow that work.

Enable Let people change the plan while it runs when people with run-control access should be able to change remaining work. Completed steps and required commitments remain protected. A plan is an agent-maintained account of progress, so read it together with the recorded work and results.

A caller can supply a plan through the SDK when the agent's plan policy permits it. See Agents and conversations for the integration. Detailed plan policy is an advanced definition setting, rather than a separate current console editor tab.

Define acceptance criteria

Open Overview → Acceptance criteria. The output-schema check is required. You can add self-assessment criteria and mark whether each is mandatory.

A useful criterion says something inspectable: “Identify missing information rather than inventing it” is more actionable than “Be excellent.” Keep the objective, output fields, and acceptance criteria consistent. A criterion requiring a source reference needs a suitable output shape and a source the agent can access.

The result must satisfy the configured checks to be accepted. Schema validation verifies structure; self-assessment is a model judgment. A successful self-assessment does not independently establish factual correctness, so include representative test cases and human review where the process needs it.

Understand milestones

A declared milestone gives the agent a named event it can report, such as research complete. Depending on the definition, it can carry a summary, structured details, or an evidence requirement. Required milestones participate in acceptance checks.

Milestone declarations and advanced evidence settings are authored through the definition API. They are not a general-purpose “Milestones” configuration screen in the current agent editor. Run history and developer interfaces expose the resulting reports.

Use milestones for meaningful progress events. Use a plan for a changing list of work. Use a workflow step when you need the system to enforce a specific transition.

Approval is a different decision

An acceptance criterion asks whether the result meets its requirements. An Approval step asks an authorized person to decide on a specific proposal. If sending a message or updating a record needs review, make that review an explicit part of the process and connect the write to it.

In a workflow, the sending or writing tool can identify its required preceding approval. Changing a proposed action after approval is not the same approved action. The workflow guide demonstrates separate approved and rejected paths.

When work does not pass

Inspect the run's outcome, acceptance details, plan, and recorded steps. Determine whether the problem is the task itself, missing capabilities, incompatible output, incomplete required work, or an overly restrictive limit. Adjust the draft, publish, and test a new version with the same input.

Do not solve an unclear acceptance failure by removing every check. Give the agent a task it can complete with the information and capabilities it actually has.