Skip to main content
A run is one execution of a mission from the first step to the last. This page covers how to start one, watch it, control it, keep it within budget, and answer the human approval gates it hits along the way.

​
Starting a run

Click Run on the mission page. To start a run you need write access — you must be the mission’s owner or a lead — and the mission must be active (not paused or archived).
Only one run can be active per mission at a time. If a run is already running, paused, or waiting for a human, you can’t start a second one — resume or finish the current one first. This is what keeps a mission’s history a clean, single timeline.

​
The run lifecycle

A run moves through a small set of statuses:

​
The live view

Runs › Live is the real-time picture of the current run: which step is running or waiting for a human, how many steps are done, and a credit-budget bar. Two controls sit here:
  • Stop after current step — a graceful stop; the running step finishes, then the run halts.
  • Cancel now — an immediate stop. The step in flight may already have had side effects (an email sent, a record written), so use it knowingly.
Past runs live under Runs › History, with a per-step Trace (prompt, answer, attempts, cost), the assembled Prompt, Tool calls, Logs, an Evaluation, and a per-step Cost table. Files produced by runs are collected under Runs › Artifacts.

​
Pause, resume, cancel

Resume is safe to press. Because it continues from the first non-success step, resuming a paused or interrupted run never re-runs completed work.

​
The credit budget

Each agent step costs exactly 1 credit. A mission can set a credit budget per run (in Settings › Budgets) — the most credits a single run may spend. When the next agent step would push the run over its budget, the run doesn’t quietly overspend. It parks as awaiting_human with a budget-exceeded status and waits.
1

A run hits its cap

The next agent step would exceed the budget, so the run parks instead of running it.
2

You're asked to raise the budget

The mission’s Overview shows the parked run with an Increase budget action.
3

Add credits and continue

Raise the budget and the run resumes from where it parked — no work is lost.

​
Human approval steps

A human step pauses the run and asks a person to decide before it can move on. It creates an approval token valid for 7 days and blocks until someone responds. Where approvals show up:
  • In-app — always. They appear in the mission’s Approvals tab and in the approver’s inbox.
  • Telegram — optional. If configured, the request is pushed to Telegram with real Approve and Reject buttons.
The decisions available on an approval: Who may decide: the step’s designated approver, the mission owner, or a lead.
Never approve on the AI’s behalf. A human gate exists precisely so a person makes the call. Approvals are recorded in the audit trail against whoever decided.

​
How a rejection loops back

Rejecting (or returning to a step) doesn’t end the mission — it sends the run back to an earlier step to try again. That loop is bounded by the same safety cap as automatic loopbacks: after 3 iterations it stops going in circles and auto-escalates to the owner as a human step, so a person decides what happens next rather than the mission retrying forever.
Resolving approvals from an AI client uses approval_list and approval_resolve; runtime control uses mission_run, mission_pause, mission_resume, mission_cancel, and mission_budget_increase. See the Missions API reference and Mission Recipes.