Autonomous coding agents are opening millions of pull requests — and most of those PRs get rejected.
Researchers studying multi-agent coding pipelines say the rejection problem starts long before review. When multiple agents work on the same codebase simultaneously, they duplicate effort, overwrite each other, and race to close the same tasks. A new paper introduces grite, an open-source coordination layer that stores a signed, append-only event log directly inside git — no central server required. In tests, the share of work that merely redid a teammate's task fell from 78% to 0%, and useful throughput more than tripled. The log also converges correctly across all agents, where a file-based tracker silently drops concurrent writes.
The value here is diagnostic as much as operational. The grite log is a mineable artifact: researchers can automatically recover specific failure modes — conflicting edits, lock starvation, redundant rediscovery, race-to-close — with full provenance. Pull request history, by contrast, hides most of this entirely.
The dataset, harness, and mining toolkit are all released publicly. Whether agent-heavy engineering teams adopt a coordination substrate like this depends on how much pain they're already absorbing from the collision tax — and most probably haven't measured it yet.