Technical guide

Multi-Agent Software Engineering Before the Pull Request

How task claims, append-only coordination events, replica convergence, and pre-PR observability expose duplicate and conflicting work among coding agents.

Direct answer

Direct answer: Multi-agent software engineering

Pull requests reveal the final code but hide much of the cost of concurrent coding agents: repeated discovery, duplicate implementation, conflicting edits, abandoned branches, lock starvation, and races to declare completion. A useful multi-agent engineering system records task claims and append-only coordination events before the pull request, allows replicas to reconcile deterministically, and measures useful throughput rather than raw agent activity.

Written for Developer-tool teams, engineering leaders, coding-agent builders, distributed-systems practitioners, and researchers studying AI-assisted software delivery.

Architecture

The operating principles

01

Coordinate before code review

Agents need a shared protocol for claiming, splitting, yielding, blocking, and completing work before their edits collide in a pull request.

02

Use append-only evidence

Immutable coordination events preserve intent and sequence, making duplicate discovery and conflicting ownership measurable after replicas reconnect.

03

Design for divergent replicas

Disconnected or delayed agents will observe different histories. The protocol must converge without silently discarding a valid write.

04

Measure useful throughput

Commits, tokens, and tasks closed reward motion. Useful throughput discounts duplicate and conflicting work and includes the cost of reconciliation.

05

Keep humans able to inspect the protocol

A git-native or otherwise ordinary event trail lets engineers review coordination failures without a proprietary observability stack.

Evidence

What supports this view

2026 arXiv preprint

Before the Pull Request

In the measured setup, duplicate or conflicting rework fell from 78% to 0%, useful throughput more than tripled, and the append-only log converged without silently dropped writes.

Inspect the source ↗
Public open-source system

grite

A git-native coordination implementation accompanying the paper and released corpus.

Inspect the source ↗
Public research corpus

grite corpus

The released artefacts support inspection of coordination traces and the measured setup.

Inspect the source ↗

Scope and limitations

Where the result applies

  • The controlled results do not establish that every coding team or agent framework will achieve the same improvement.
  • Coordination overhead can exceed its value for very small, independent, or short-lived tasks.
  • A convergent event log does not make the underlying code correct; testing and review remain separate obligations.

Common questions

Answers in plain language

Why are pull requests insufficient for observing coding-agent work?+

They capture proposed code after substantial work has happened. They usually omit repeated discovery, abandoned attempts, duplicate task ownership, and coordination messages that explain why the work became expensive.

What is a task claim in multi-agent coding?+

A task claim is an explicit, time-bounded coordination event stating that an agent intends to work on a defined scope. It reduces accidental duplication without requiring a permanent central lock.

What should multi-agent engineering teams measure?+

Useful completed work, duplicate and conflicting effort, claim wait time, reconciliation cost, abandoned work, human interventions, and whether every valid coordination event survives convergence.

Does more agent activity mean more engineering throughput?+

No. Parallel activity can increase duplication and conflict. Throughput should count useful, integrated work after subtracting the effort needed to resolve overlap and errors.

Continue exploring

Apply the architecture to a real system.

Bring a specific outcome, failure mode, or audience. The next step can be an architecture review, prototype, talk, or workshop.