Starting today, AI Investigation is live for all users. The moment a test fails, an agent jumps in to diagnose the issue and take the appropriate action.
- If it's a flake, the agent clears up whatever in the environment caused it and re-runs the test.
- If it's a broken test, the agent repairs the test while keeping the same assertions and re-runs it.
- If it's a bug, the agent reproduces the failure in its own browser and generates a bug report with a video playback, traces, and repro steps.
Every failure gets an answer 6.5x faster because you're not waiting for the whole suite to finish before investigating the failures. And 75% of failures are handled with no human involvement.
How it works
The first agent on a failure becomes the lead, and every later failure goes through it. A failure with a known cause goes to the investigator already working on it, and a new cause gets a new investigator with its own sandbox, browser, and branch. By making sure that one agent handles all related failures at the same time, we prevent agents from stepping on each other during a run, and one fix can clear multiple tests at once.
The run page shows each finding and the tests it affects: what's fixed, what's reported, and what's waiting on you. You can dispute any conclusion, follow along from ChatGPT or Claude through QA Wolf MCP, and give investigators instructions for your suite with skills in your workspace's .skills/ folder.