1/ The features customers use most are often the ones with no test. We're building AI agents that hunt for them 24/7, write the missing test, and send any bug it catches straight to an AI fix agent. No person in the search.
2
30
5/ Then the loop closes. The AI tester runs the new test in a real browser; if it fails, the bug goes to our 24/7 AI fix agent, which drafts and tests the fix. People review the test and approve the fix, and that's the only human step.
1
6
1/ 4,700 test steps, zero selectors, zero assertions — just prose written for human testers ("Go to the Accounts tab"). An AI agent now runs them in a real browser. The surprise isn't the cost — it's the quality.
4/ The genuinely new part is judgment: comparing what's on screen against a prose expected result, written for a human reader. One-time cost up front — teaching the agent what each test is really checking. After that, every run is the expert run.
1
21
1/ 200 of 215 API operations in one of our products — 93% of its surface — had zero production calls. Nobody knew. Old products carry far more dead code than anyone guesses, and we just started measuring it.
1
18
5/ Design choice: detection is deterministic end to end — no LLM decides what's dead, so every finding is reproducible. The AI comes in downstream: each finding is filed as a ticket that flows into our automated fix pipeline and closes itself once the code is gone.
1
6
1/ We added a production-log agent to our AI 24/7 Quality Loop, the pipelines that watch a legacy enterprise product around the clock.
It traces each recurring error to the code that emits it and judges it like a senior engineer would.
First live run: 8 real defects.
1
35
4/ What changed: the agent must put each finding in one of six classes before anything is filed, and deduplication runs in plain code, not in the model.
In the first run, 70% of signals were set aside, each with a written reason.
1
4


