User guideAI Engineering Platform
Back to siteSite

Agent runs and approvals

A run is one agent execution. It has a type, a frozen context, a budget and an event stream. The run view is where a team sees what the agent did and decides what happens next.

A completed run

A development run: plan, approval, build, test and review in one timeline, with the diff that will be merged.

Run types

TypePurposeGate
PlanningInspect context and propose steps, files and risksPlan approval
DevelopmentImplement the approved plan and produce a diffTest/design gates, PR approval
ReviewInspect a diff and report findingsPR approval

Start a run

  1. Open a Ready task.
  2. Choose Run agent and select the type.
  3. Confirm the model/agent and budget if the project allows a choice.
  4. Watch the run view; events stream live.

A run cannot start if the task is not ready, if the repository is unreachable, or if the project has no allowed agent.

What the run view shows

  • Context — the frozen manifest: files, rules and token budget
  • Plan — steps, files, risks and estimated cost
  • Approval — the decision and who made it
  • Changes — the diff, with additions and deletions
  • Checks — tests, design checks and acceptance evidence
  • Cost — recorded tokens and provider cost
  • Events — an append-only timeline with timestamps

Approvals

Plan approval

A human can approve, request changes or reject.

  • Approve sends the run to implementation.
  • Request changes returns it to planning with a note.
  • Reject stops the run and leaves the task for a human.

The approval records the user, the time and the plan version.

PR approval

A pull request requires the configured number of distinct approvers and all required checks. An agent cannot approve its own PR or bypass branch protection.

Retry, cancel and failures

  • Retry starts a new run from a checkpoint or from the beginning, depending on the failure.
  • Cancel is cooperative: the worker asks the runtime to stop and kills the sandbox.
  • A failed run keeps its events and artifacts for inspection. The task returns to a state a human can act on.

Common failure reasons are shown in the run view: clone failure, provider error, budget exceeded, missing proof or a failed check.

Budgets

A run has a token and cost budget. The worker checks it between phases. When the budget is exhausted the run fails with a clear reason instead of continuing silently.

Next

Retry development, Fix review findings and Fix CI

The run page offers a different action depending on where the run stopped:

The actionWhen to use itWhat it does
Retry developmentthe development run failed (a build, a test, a timeout)re-queues the same task from the plan; a human still approves the new run
Retry taskany failed runre-queues the task from the start
Fix review findingsthe review run found blockersstarts a fix run that addresses the findings and re-runs the gates
Fix CIthe pull request's checks failedstarts a fix run that targets the failing check

Every one of them is a new run with its own trace; the platform never merges.