I built an open-source multi-agent SDLC harness that beats a cold Claude Code run on large repos

A developer released an open-source multi-agent harness for the SDLC that outperforms Claude Code on large repositories by caching repo structure knowledge in a single learning pass, with published benchmarks.

This shifts the cost model for codebase automation from per-query context loading to one-time upfront indexing, making multi-agent orchestration economically viable for large, unfamiliar repos. For operators, it removes the bottleneck of recalculating structural understanding across every agent call, which currently limits cold-start performance of single-agent tools like Claude Code. Second-order effect: teams may adopt a standard “repo pre-index” step before any agent-driven work, creating demand for persistent knowledge stores tied to codebases rather than ephemeral sessions. The approach also reduces the advantage of proprietary models that rely on built-in context windows, lowering infrastructure lock-in for large-scale code tasks.