Autonomous coding agents now open millions of pull requests, yet large-scale studies find their PRs are produced faster but accepted less often — a coordination and trust gap that pull-request-level telemetry cannot explain. According to a paper by Sarkar and Dipankar (arXiv:2606.19616), the missing signal lives before the PR, in how concurrent agents claim, divide, and collide over shared work.
The researchers introduce grite, an open-source coordination substrate that needs no central server and stores its records inside git itself. Its append-only, signed event log captures the coordination process directly. The study demonstrates that this shared substrate reduces duplicate and conflicting work at bounded overhead. Specifically:
| Metric | Before grite | After grite |
|---|---|---|
| Duplicate/rework share | 78% | 0% |
| Useful throughput | Baseline | More than triples |
| Log consistency | File-based tracker loses writes | Converges, no silent drops |
Reducing Duplicate and Conflicting Work
The share of work that merely re-does a teammate's task fell from 78% to 0%, while useful throughput more than tripled. Every agent's copy of the log converges to the same state with no write silently dropped, where a file-based tracker loses concurrent writes. These results, the paper states, are achieved at bounded overhead.
Mining the Coordination Log
The log is a mineable artefact from which concrete failure modes — conflicting edits, lock starvation, redundant rediscovery, race-to-close — are automatically recoverable with provenance. Several of these failure modes are invisible in pull-request history. The researchers release the dataset, harness, and mining toolkit alongside the paper.
Implications for Enterprise AI Agent Workflows
For enterprises deploying multiple AI coding agents, the findings suggest that monitoring coordination before the PR stage can dramatically improve acceptance rates and reduce wasted work. The approach requires no central server and lives inside git, making it lightweight to adopt. The ability to automatically recover failure modes with provenance could help engineering teams debug agent interactions that currently go undetected.
The study, published on arXiv under the Computer Science > Software Engineering category, is available with full code and data. The authors advocate for moving beyond PR-level metrics to understand the full lifecycle of multi-agent software development.
