AI Autopilot: Flow of accepted tasks grew thirteenfold
Anthropic
OpenAI
An engineering team integrated an AI autopilot pipeline into their workflow, from YouTrack tickets to verification in a live GUI. The number of verified tasks jumped from under 5 per week to about 60, with AI handling three-quarters. The system also cleared a large backlog, but reopen rates rose initially, prompting the addition of an independent audit step.
The team built an AI autopilot pipeline that takes tasks from YouTrack through to verification in a live 2D/3D GUI. Before the autopilot, the project accepted fewer than 5 tasks per week; after, about 60, with the AI handling three-quarters. New ticket submissions also grew nearly threefold, partly because users started reporting issues now that fixes were fast. The median time to first fix dropped from nearly three days to less than one. The team cleared a backlog of over 700 tasks, taking 144 aged over nine months, and nearly two-thirds were accepted. However, the reopen rate rose from about 11% to higher levels, often because AI-generated proofs were unconvincing. In response, they introduced an independent audit pass, which rejected work in about one-third of runs and caught misinterpreted tasks. The system ran 869 hours across nearly eight weeks, equivalent to roughly 2.8 full-time engineers. The monthly subscription cost was about $500, while equivalent token-based pricing would be about $5700. Claude was cheaper but less strict, while Codex was more expensive but better at GUI verification. The team plans to move Codex to a GUI pilot role and use a simpler model for auditing.
- Сокращения
- GUI = Graphical User Interface — графический интерфейс пользователя
Source: Habr — хаб ИИ —
original
