An agent says the fix is done. The test is green. The question is what that green result proves. 9lives runs your existing Playwright tests and records the execution evidence so you can inspect the answer.
Our new launch film introduces the workflow in 60 seconds. The interactive demo goes further: a missing assertion, a recorded agent session, and a patch checked against unfixed and fixed app revisions.
Watch the 9lives launch film
Agent fixes. 9l checks the change.
The Prove the fix chapter follows a profile-form bug. The agent writes a patch. 9l then runs the same reproduction test bytes against separate checkouts of the application: the assertion fails before the patch and passes after it.
That is the role of experimental fix confirmation. The agent authors the change; 9l executes the comparison and returns attributable receipts. A confirmed result establishes that this reproduction distinguishes the two revisions. The assertion still needs to check the reported defect.
Next, the recorded run injects three faults into a profile request. The test fails on an assertion for all three. That gives us evidence that the test notices those faults. It does not establish complete coverage or correct handling of every error; the HTTP 401 case is synthetic, not a real login flow.
What you are seeing: The demo separates interactive browser walkthroughs, an algorithm walkthrough, and recorded execution. Its real agent session includes scrollable logs and MCP calls. Opening a chapter does not start a fresh agent run. Fix confirmation and fault testing remain experimental; these results qualify one synthetic app in Chromium.
A green test can still miss the wrong target
Two rows can both have an “Edit” button. Clicking the wrong one can still open a visible editor, so an assertion that only checks visibility can pass while the action targets the wrong record.
Our separate duplicate-target fixture makes that failure observable. An exact record-ID assertion catches the wrong row. A stable-ID locator passes after the rows are reordered, and a deliberately wrong expected ID fails. The receipts show which assertions actually ran. This fixture demonstrates target selection and assertion quality; it does not demonstrate automatic healing.
Keep your Playwright investment
The Go runner works with your existing Playwright specs and configuration. Go owns execution budgets, attempt identity, cancellation, and completeness accounting. Playwright owns the browser and executes your assertions.
A passing receipt means the assertions passed and the required execution evidence was complete. Missing or malformed evidence cannot turn an incomplete run into a pass. It does not mean the application is defect-free or every requirement is covered.
The optional @9l/playwright SDK adds steps and bounded natural-language goals. Keep explicit assertions after those interactions: goal completion alone is unverified. You can also connect an agent through MCP to inspect a page, execute a reproduction, and request fix confirmation.
The Behind the agent chapter explains the contracts between the agent, 9l, and Playwright MCP. Observed controls and bounded actions make the execution inspectable; the final assertion establishes the tested outcome.
Try one spec on your machine
Native CLI binaries are available for Linux, macOS, and Windows, on x64 and ARM64. No QualityMax account is needed for the local runner. Your project still supplies its Node dependencies and Playwright browsers.
Follow the existing-project guide: inspect the planned tests, run a small set with a time limit, and inspect the receipt. Include a known failure as well as a happy path. Check that the failure comes from the assertion you intended.
Use 9l assess for advisory static findings about test patterns. Static assessment does not execute assertions or establish coverage. When exploring repairs, review the candidate diff and the completed re-run before accepting the change. Native selector healing is experimental; the Go 9l heal command still delegates to Python.
The original field trial: a historical Python result
October 10–11, 2026: The repair and AI-goal observations below came from the older Python 9lives prototype, version 0.3.0, alongside a healthcare suite with 88 spec files and roughly 900 Playwright tests. The focused runs exercised selected cases, not the full suite. They are not timings for the current Go runner.
- Renamed button · offline repair
- 20 seconds · 0 model calls · 7/7 passed on re-run · candidate saved for review.
- Drifted field ID · Claude-assisted repair
- 39 seconds · 1 Claude call through an existing subscription · candidate verified.
- Failure inside an assertion · human review
needs_human· 0 model calls, with and without a configured provider.
In those cases, repair changed supported locator literals and saved a .healed candidate for review. When drift appeared inside an expect(...).toBeDisabled() assertion, it stopped for human judgment. Changing that check could change what the test proves.
Two login-page goals completed in 5.7 and 7.5 seconds through a signed-in Claude CLI, and explicit assertions after the goals passed. A goal requesting a nonexistent button returned unresolved and the test failed. These observations describe that historical environment; provider subscription access remains subject to the provider’s limits.
We also inspected TesterArmy e2e 0.19.0 and ran a small deterministic comparison: four tests took 1.5 seconds in e2e and 1.7 seconds in Playwright. We did not run e2e’s AI steps or AI repair, so this trial cannot establish comparative healing speed. It was a small field trial, not a general benchmark.
A local companion to QualityMax
Use 9lives when the work starts in your repository: execute a test, assess its checks, inspect a receipt, or compare a reproduction across a patch. Use the QualityMax platform for the team’s broader workflow of discovery, generation, managed execution, and review.
QualityMax and 9lives were built without external funding. We want the results to speak through runs developers can inspect and reproduce.
Start with one assertion you trust.
Watch the film, inspect the demo, then try a bounded run on your own Playwright project.
Inspect the proof →