Open-source SDLC agent harness says it beats cold Claude Code on large repos

NeighborhoodOwn8510 · reddit · 2026-07-24

An open-source project called AutoDev Studio claims to outperform a cold Claude Code run on large repositories by paying the repo-localization cost once and reusing it across tasks.

What it does

Ingests a Git repo and builds a durable knowledge base from static analysis plus a local embedding index.
Runs a multi-agent SDLC pipeline: PM agent for clarification and tickets, Dev agent for code changes, QA/testing, and a separate reviewer model.
Opens a real PR and tracks token and dollar costs per task and per agent.

Results reported

On textual (82k LOC) and rich (35k LOC), the tuned pipeline beat a cold claude -p on 6/6 well-localized tasks.
Savings ranged from 7% to 75%.
In one bug, the cold agent spent $6.83 and 207 turns locating the issue, while the pipeline fixed it for about $1.70.

Limits and caveats

On tiny, trivially greppable edits, the fixed five-stage overhead can cost more than it saves.
On one cross-cutting bug, it produced a cheaper but narrower fix than the baseline.
The stack is provider-agnostic, supports Anthropic, Claude Code CLI, and OpenAI-compatible endpoints, and can run with a free local embedding model.

Original post →

Open-source SDLC agent harness says it beats cold Claude Code on large repos

More from coding & agent