arXiv
Before the Pull Request: Mining Multi-Agent Coordination
Dipankar Sarkar
Abstract
Multi-agent coding is the new inner loop. Five agents converging on a single correct solution is the new test of orchestration. We mine the coordination patterns that precede successful pull requests, and we show that the agent that gets the right answer is rarely the agent that gets the credit. The framework is implemented in grite, an open-source toolkit (github.com/neul-labs/grite).
The new inner loop of software engineering is multi-agent coding. Five agents converging on a single correct solution is the new test of orchestration.
Abstract
We mine the coordination patterns that precede successful pull requests, and we show that the agent that gets the right answer is rarely the agent that gets the credit. The framework is implemented in grite, an open-source toolkit (github.com/neul-labs/grite). Published June 2026 on arXiv (2606.19616).
Frequently Asked Questions
What is multi-agent coordination in software engineering?
Multi-agent coordination is the practice of having multiple AI agents collaborate on a single coding task. The agents divide the work, share context, and converge on a solution. This is becoming the new inner loop of software engineering: instead of one developer, you have five agents working in parallel. We study the coordination patterns that lead to successful pull requests in this paper (arXiv:2606.19616, June 2026).
What is grite and where can I use it?
grite is the open-source toolkit that implements the multi-agent coordination framework from this paper. It is at github.com/neul-labs/grite. grite provides the coordination primitives (handoff, consensus, dissent), the metrics (time-to-PR, consensus rate, agent-credit), and the evaluation harness. It is built to work with any LLM agent framework (LangGraph, AutoGen, CrewAI).
What does 'the agent that gets the right answer is rarely the agent that gets the credit' mean?
Our mining shows that the agent that produces the final accepted PR is rarely the agent that produced the first correct solution. The credit attribution is biased toward the agent that pushes the PR through review, not the agent that first saw the right answer. This has implications for how you structure incentives in multi-agent systems: reward the right answer, not the loudest voice.
Why is multi-agent coordination the new test of orchestration?
Single-agent coding is a commodity — the model is the same, the prompt is the only differentiator. The differentiation in 2026 and beyond is in the orchestration: how do you split work across agents, share context, and converge on a solution. Multi-agent coordination is the orchestration question. This paper presents the empirical study of what works.