Outcomes and safety limits

A Run can finish every step and still not do the job. The Outcome tab defines what "done" means. The Safety tab stops a Run before it goes too far.

Outcome: define success

On the Outcome tab, under What counts as done, you can set:

1. Results that must be present — named results, separated by commas. Any step can produce them, and a recorded action counts as proof.
2. Results that must have a particular value — values those results must have.
3. In your own words (optional) — a plain-language check, such as "The summary contains exactly one next action." The plain-language check is evaluated by AI against the Run. Treat its verdict as a signal to review—not proof that the task succeeded or that its facts are correct.

Click Save definition of done when you finish.

The verdict looks at the whole Run: every step's output and every external action, not only the last step. A Run whose steps all finished can still fail its outcome.

Reading the verdict

• The verdict is a second opinion, not a status change. A Run that finished every step stays completed even when the verdict says something was not met. Read the Run's output yourself and decide.
• If no verdict appears, refresh first. The verdict can take a moment after a Run finishes. Refresh the Run before you conclude it is missing.
• The checks use shortened evidence. Outcome checks can look at a shortened version of a step's output. If a verdict says text is missing or cut off, expand the step and compare its full output before you retry or change the workflow. If the verdict still disagrees with what you can see, ask your admin or contact support.

Keep it simple and measurable

Good outcomes can be checked without guessing:

Required result: summary
Required result: next_action
In your own words: The summary only contains facts stated in the run objective.

Avoid vague goals such as "the summary is good." A check nobody can measure gives you a verdict nobody can trust.

For your first Loop

The first Loop returns no named results, so it uses only In your own words — a plain description of what a good summary contains. That check is judged by AI, so it will not pass every Run automatically. Treat a not-met verdict as a reason to read the output yourself, not as proof the Loop is broken.

Safety: set limits

Every Run has four limits. If a Run reaches one, it stops.

To change them:

1. Open the Safety tab.
2. Enter new values.
3. Click Save limits.
4. Publish the Loop. New limits apply to future Runs.

Choose your own limits

The defaults are generous. Lower limits catch problems sooner.

1. Run several tests with realistic inputs.
2. Open each Run and note the steps, tokens, and time it used.
3. Set each limit a little above the highest value you saw.
4. Set external actions to the number your Loop actually needs. Use at least 1, even for a text-only Loop.

For the two-step first Loop, the tutorial uses 4 steps, 10,000 tokens, 300 seconds, and 1 external action.