Open-source AutoDev Studio says it can fix localized bugs in Python repos for up to 75% less
NeighborhoodOwn8510 · reddit · 2026-07-27
An open-source SDLC harness called AutoDev Studio claims it can beat cold Claude Code runs on cost whenever a task can be localized well.
The author says a cold Claude Code run spent $6.83 and 207 turns chasing one bug in an 82,000-line repo, while AutoDev Studio fixed the same issue for about $1.70. The workflow is a staged software lifecycle:
- a PM agent clarifies the request and drafts concrete tickets
- a human approves before any code changes
- a Dev agent works on an isolated branch
- QA runs the repo’s real tests
- a reviewer from a different model family checks the diff
- bounded revise loops handle failures conservatively
- the system opens a real PR for human merge
The post emphasizes that every stage records real tokens, cost, and duration, and that different stages can use different providers and model families. It can run on existing subscriptions for Claude Code, Codex, Cursor, Aider, or Gemini CLI without juggling API keys. Benchmarks across two large Python repos found 6/6 localized tasks were 7% to 75% cheaper, though trivial one-line edits can still cost more and one cross-cutting bug got a narrower fix than the baseline.
More from coding & agent
- Open-source SDLC harness beats Claude Code by up to 75% on cost in repo-localized tasks — NeighborhoodOwn8510 · 2026-07-27
- PrimeTask adds a local MCP server for ChatGPT and Claude to run work apps by chat — XVX109 · 2026-07-27
- Hermes power users get better results with memory, skills, and cron — tomcrawshaw01 · 2026-07-27
- A font-design skill has been built and open-sourced — JeremyNguyenPhD · 2026-07-27
- LLM skills can speed teams up, but they also add technical debt — srchvrs · 2026-07-27
- Annotate lets AI agents draw arrows and labels on your screen via MCP — adamm_carter · 2026-07-27