This Open-Source Tool Cuts Duplicate AI Agent Work from 78% to 0% — And It's a Game Changer for Coordination
Autonomous coding agents open millions of PRs but get rejected—grite fixes the root coordination gap.
A new study by Dipankar Sarkar tackles a growing problem in AI-assisted software development: autonomous coding agents generate millions of pull requests, but large-scale analyses show these PRs are accepted less often than human ones. The missing signal, Sarkar argues, lies before the PR—in how concurrent agents claim, divide, and collide over shared work. To capture this, he built grite, an open-source coordination substrate that needs no central server and stores its append-only, signed event log directly inside git. This design ensures every agent's log converges to the same state with no writes silently dropped, unlike file-based trackers.
The results are striking: with grite, the share of work that merely re-does a teammate's task fell from 78% to 0%, while useful throughput more than tripled. The log also makes concrete failure modes—conflicting edits, lock starvation, redundant rediscovery, and race-to-close—automatically recoverable with provenance, several of which are invisible in pull-request history. Sarkar released the dataset, harness, and mining toolkit for further research. As AI agents increasingly collaborate on code, tools like grite address a critical bottleneck: not just writing code faster, but coordinating reliably.
- Duplicate work dropped from 78% to 0% after implementing grite's coordination substrate.
- Useful throughput more than tripled, with no silently dropped concurrent writes.
- Failure modes like conflicting edits and lock starvation are automatically recoverable from grite's git-based log.
Why It Matters
As AI coding agents proliferate, coordination tools like grite are essential for reliable, efficient collaborative development.