Auto-improve is a second AI that studies the first. It reads back through your Agent's runs, finds what's going wrong, and suggests changes to its instructions.
Go to Agent → open an Agent → Improve → Auto-improve. The page has two tabs: Configure and Analyses.
How do I run an analysis?
Open the Analyses tab, tell it what to look at if you want, and click Run analysis.
Look at the runs where the customer asked about refunds and we ended up escalating.Leave the box empty and it looks at everything. When it finishes, it tells you how many runs it read, how many problems it found, how many fixes it made, and what it cost.
Where do I find past analyses?
In the table on the Analyses tab, newest first. Each row is one finished analysis:
| Column | What it shows |
|---|---|
| Started | When the analysis began |
| Trigger | Scheduled, or Manual if someone clicked Run analysis |
| Status | Completed, Failed, or Canceled |
| Issues found | How many new issues it opened |
| Spend | The credits it used |
| Duration | How long it took |
Click Started, Issues found, Spend, or Duration to sort by that column. Use Filter to show only one trigger or status. Click a row to open the analysis and see the runs it read and the issues it found.
The table lists your 100 most recent analyses.
How do I tell it what "good" looks like?
Open the Configure tab and fill in three fields:
| Field | What to write |
|---|---|
| Goal | The goal, in your own words. Resolve billing questions in one reply, and never guess at an invoice amount. |
| Optimize for | Pick any of Quality, Cost, Duration, Tool accuracy, and Safety |
| Additional context | Things that look like mistakes but aren't. Refunds over $500 go to a human. That is policy, not a failure. |
Skip that last one and Auto-improve will report your own policies as bugs.
Can it run on a schedule?
Yes. On the Configure tab, turn the schedule on, set when, and pick a timezone.
Can it fix things on its own?
Yes. Under Behavior, switch it from only raising problems to fixing them automatically. Leave it on raising problems until you've read a few analyses and trust what it suggests.
How do I stop it spending too much?
Three limits sit under Behavior and Limits on the Configure tab:
| Limit | What it limits |
|---|---|
| Runs per analysis | How many of your runs one analysis reads |
| Credit limit | What one analysis can spend before it pauses |
| Time limit | How long one analysis can go on before it pauses |
Every analysis reads some runs, so Runs per analysis is required: enter a whole number from 1 to 50. Leave Credit limit or Time limit empty for no limit.
Auto-improve costs credits, but it isn't a run, so it shows up under Spend by source on Analytics rather than in the run charts.
Where do its findings go?
To the Issues page. The Analyses tab shows the analyses it's done. Issues shows what they found.
Additional FAQ
What does it read?
Your Agent's runs, including the Review your team set on them (Approved or Rejected) and the notes saying why. See Run history. Reviewing runs makes its analyses better.
Can the Agent see its own issues?
Only if you let it. Manage → Agent settings → Issue access has three settings: None, Read, and Read and write. See Issues.
Does an analysis change my Agent?
Not unless you apply a fix, or you turned on automatic fixes. Applying one is immediate: it publishes a new version and goes live on the next run, with no draft to check first. Every fix is in your version history, so you can roll it back. See Issues.
How often should it run?
Often enough that there are new runs to learn from. For a busy Agent, once a week is a good place to start.