Behind every success are many times as many failures. Here's every way AutoDev LG has stumbled while building an app, counted by kind. It updates itself every day from the run records.
Public tests: tests the worker writes to check its own code. Acceptance tests: tests written by the test writer that give the final verdict on whether the app matches the spec.
At the test-writing stage, the pass/fail tests (Acceptance tests) couldn't be written correctly, so it ended before any coding.
"Can't build the app before the answer key exists."
Episode: In inv18 on 9/28, Qwen3.8-27B, then Gemma-4-26B-A4B-it, then gpt-oss-120b each took a turn — and all three dropped out here.
The app got built, but failed the Acceptance tests that judge whether it matches the spec. It got as far as diagnosing why, then stopped.
"The type that trips right before the finish line."
The Public tests the worker wrote to check its own code didn't pass. It diagnosed why, then stopped.
"Sets its own exam, fails its own exam."
Episode: In inv01 on 9/26, a screen test the worker wrote crashed inside the test itself.
The run itself stopped — a timeout, or the workload hit its cap.
"Ran out of steam."
Episode: In inv16 on 9/27, a test run timed out after 180 seconds. inv22 on 9/29 ran for about 2 h 8 min before hitting OpenHands' workload cap.
It tried to repair the app, but decided the fix couldn't be done within the files it was allowed to touch.
"It knows the problem is somewhere else."
It tried to repair the failing tests, but made no progress and gave up.
"The more it fixed, the deeper it sank."
It tried to rewrite the Acceptance tests, but the check rejected the new ones.
"No do-overs on the do-over."
In the very first step, the requirements couldn't be matched to the spec's items, so design never started.
"Tripped on the starting line."
The tests passed, but the final check of the finished app against the spec found a mismatch.
"Perfect score on the tests — forgot to write its name on the paper."
A proposed repair to the harness that runs the Acceptance tests was rejected by the check.
"The repair quote didn't get approved."
Counted from the same run records as the dev log (stopped runs excluded).