# Give your agent a program, not a mouse

> After a person or an AI agent does a data-entry task, OpenAdapt turns that run into a program that makes the same entry again without calling an AI model. The program confirms it has the right record before it saves and reads the record back afterward. When something doesn't match, it stops and asks a person.

- Canonical: https://openadapt.ai/
- Updated: 2026-11-21

## Overview

Your agent, or a person, does a task once in an app. OpenAdapt compiles it into a program your agent can call. A healthy run makes no model calls; the program checks the record before and after each save, and stops when a check fails.

## Watch it catch a bad save on your computer

In about a minute, it runs a clinic app with synthetic records on your computer, makes one entry that saves and one whose save fails on purpose, and shows OpenAdapt stopping. No account, and nothing is uploaded.

```bash
uvx --python 3.12 openadapt@latest quickstart --break-it
```

- [Inspect the evidence](https://app.openadapt.ai/demo)

## It reads the saved record back

On a test copy of OpenEMR, Claude Sonnet 5.5 entered referrals on its own with computer use while we made some of the saves fail in the database. The agent reported 28 of the 30 failed saves as done. Reading the records back from the database caught all 30, and it confirmed the 5 that saved correctly. The table shows each kind of failure.

- **28 of 30** — bad saves the agent reported as done
- **30 of 30** — injected bad saves caught by reading the record back
- **0** — false alarms on clean saves
- **$0** — in model fees per run. The compiled program calls no model.

| Case | The agent said | The read-back found |
| --- | --- | --- |
| A field blanked | 4 of 5 done | 5 of 5 caught |
| A value changed | 5 of 5 done | 5 of 5 caught |
| A duplicate chart | 4 of 5 done | 5 of 5 caught |
| The save never landed | 5 of 5 done | 5 of 5 caught |
| Another chart deleted | 5 of 5 done | 5 of 5 caught |
| Another chart edited | 5 of 5 done | 5 of 5 caught |
| Clean saves | 5 of 5 done | 5 of 5 confirmed |

The agent noticed none of the 30 failures. On the other 2 it also said it saved the chart, but added that it couldn’t see the whole email field. On a second run with 30 different injected failures, the read-back caught all 30 injected bad saves plus 1 bad save the agent made itself on a clean-save item (31 of 31 in all). One item in each run where the saved chart didn’t hold the email the agent typed, so the fault never fired, was also not reported as verified and is outside those denominators.

## Show it once. Run it every time.

One harness, three modes. A model works only while a task is shown or repaired, on your own key.

- **Explore — Show the task once**: Your agent or a person does the task while OpenAdapt records every action. The save is read back before anything compiles.
- **Run — The program repeats it**: New inputs, the same compiled steps, no model calls. It confirms the record before each save and reads it back after.
- **Repair — When the app changes**: The run stops before saving. An agent shows the fix, the program tries it up to the save without saving, and a person approves it.

## Your agent supplies the values, and the program enters them in the app

Your agent calls a compiled program by name. The program enters the values, reads the record back and returns the outcome.

### Claude Code: one command

```bash
claude mcp add openadapt -- uvx --python 3.12 openadapt-agent serve
```

Adds OpenAdapt to Claude Code. With no workflows of your own, it serves a sandbox workflow.

## OpenAdapt runs on a computer you control

It signs in to your system from there, and your records stay there. It suits work where a wrong entry is costly, such as a clinic’s charts or an insurer’s claims.

- **A person decides each save**: By default, each save waits for a person, with a picture of exactly what will be saved.
- **The record is read back**: After a save, OpenAdapt checks the saved fields through a separate read-only login.
- **The run report is signed**: Anyone can check its signature in a browser.

## Our test runs and results, with their evidence

Our test runs use test systems and synthetic records, and each is published with its evidence. Your own runs stay private.

- [Receipts](https://openadapt.ai/receipts): Check each run’s receipt and its signature in your browser.
- [Tested runs](https://openadapt.ai/workflows): Systems and tasks, with dated results.
- [Processes](https://openadapt.ai/templates): Referral intake, claims, eligibility and more.

## Sources

1. OpenAdapt scoreboard, the agent's claims versus the read-back, run openemr_referral/2026-11-20-a ([SCOREBOARD.md](https://github.com/OpenAdaptAI/OpenAdapt/blob/main/packages/flow/benchmark/scoreboard/SCOREBOARD.md#the-agents-claims-versus-the-read-back)). Claude Sonnet 5.5 did the entry alone with computer use, on a local OpenEMR test copy with synthetic patients. We injected faults only in the test database, in these ways: a field blanked, a value changed, a duplicate chart, a save that never landed, another chart deleted, and another chart edited. Each fault had 5 items, and there were 5 clean items as well. By the scoreboard's success rule, the agent reported 28 of the 30 bad saves as done; on the other 2 it said it saved the chart but couldn't see the whole email field. It noticed none of the 30 failures. verify-run (openadapt-flow 3.3.0) caught 30 of the 30 injected bad saves. It confirmed 5 of 5 clean saves, with 0 false alarms. One more item, where the saved chart didn't hold the email the agent typed so the fault never fired, was also not reported as verified. A second run, openemr_referral/heldout-2026-11-20-a, used 6 different faults with three existing charts: verify-run caught 30 of 30 injected bad saves, plus 1 bad save the agent made itself on a clean-save item (31 of 31 in all), and confirmed 5 of 5 clean saves with 0 false alarms; one fault-not-reached item was also not reported as verified; in that run the agent itself reported the wrong date of birth in all 5 items where the fault moved it. The runs weren't timed.
2. OpenAdapt quickstart with --break-it (openadapt launcher 1.17.0), MockMed synthetic clinic app, 18 Oct 2026 ([the run's reports](https://github.com/OpenAdaptAI/openadapt-web/tree/main/data/evidence/quickstart-break-it-2026-10-18)): the normal entry ended Done and checked by a read of the saved note; with the save rejected on purpose, the screen looked the same as the good save, and the run stopped and reported Check the record without retrying.
