Multi-agent coordination framework boosts AI software engineering accuracy by up to 26%

AI agents have made significant strides in tackling individual software engineering challenges, such as fixing bugs and resolving GitHub issues. However, they struggle with larger projects that require coordinating work across many interconnected subtasks and completing them efficiently. Running multiple agents at the same time should help speed things up, but this approach introduces new problems—concurrent changes collide, task dependencies become hard to manage, and merging partial results into a working whole becomes unwieldy.

A Coordination Strategy Inspired by Human Development#

Researchers have drawn inspiration from how human development teams tackle these challenges in real codebases. The new approach, called Centralized Asynchronous Isolated Delegation (CAID), adapts three time-tested practices from software engineering: assigning work from a central point, letting team members proceed in parallel, and giving each worker their own space to avoid stepping on each other's toes.

In CAID, a central coordinator builds a task plan that accounts for dependencies between subtasks. Multiple agents then execute their work concurrently in isolated workspaces. Progress gets merged back together and validated through automated testing, ensuring changes fit together coherently.

Measured Improvements Across Real Tasks#

Testing revealed notable gains. On paper reproduction work (tracked via PaperBench), CAID delivered a 25.6% absolute improvement over single-agent approaches. For tasks involving Python library development (Commit0), the accuracy boost reached 14.7%.

Why Git Primitives Matter#

Analysis of the results highlights that branch-and-merge is a critical component for coordinating multi-agent work. Standard git tools—particularly git worktree, git commit, and git merge—provide the reliable, executable foundation that makes this coordination model practical and robust.

The research draws a direct line between mature collaboration infrastructure that human teams use every day and the mechanisms needed for AI agents to work together effectively on complex software tasks.