Swapping Claude for Gemini mid-run at step 433: disk-based init-d state files make agent handoff seamless

RandalSchwartz · reddit · 2026-10-23

A Reddit dev shares a multi-step agent coding workflow inspired by a Harvard study of 718 companies: AI-assisted devs write 30% more code but close zero extra tickets as PR review waits jumped 49%.

His fix: move all agent state out of the chat window into a Makefile and numbered Unix init-d state files on disk (00governance through 99nextaction), plus a read-only Codebase Archaeologist that checks git history before code is promised, and a local Critic grading diffs before git commit. Five production tickets (926 steps) ran with zero human nudges; the Archaeologist once killed a ticket in 8 minutes by proving the requested auth wrapper was already built and rejected two weeks prior.

Most striking: when Claude Sonnet hit a quota wall at step 432 of a 960-step run, switching to Gemini let it read the on-disk state files and finish on the first try. In contrast, open-weight GPT-OSS 120B in the same cockpit ignored the state files, hit 6 context resets, searched Google 13 times for private code, and fork-bombed git commit with a 0ms timeout while rm-ing the index lock.

Original post →

More from AGI Musings

AGI Musings channel →