Coordinating Agents on a Shared Git Workspace: An Empirical Study of Action-Space Observation for Agentic Software Engineering

Abstract

Background. Multi-agent approaches to Large Language Model (LLM)-based agentic software engineering, specifically in automated program repair and issue resolution, have converged on architectural patterns that favour isolation: role specialisation, task decomposition in isolated worktrees, or best-of-N with external selection. Collaboration of homogeneous agents on a single task in a shared Git workspace remains understudied, and is actively discouraged in production guidance due to file-level write collisions. 

Aims. We investigate whether a shared-workspace configuration can be made viable and effective for autonomous bug fixing and patch production through action-space coordination - a paradigm in which agents observe one another’s commits, lifecycle transitions, and recent test or error outputs rather than exchanging natural-language messages. 
Method. We propose PASC (Peer-aware Action-Space Coordination), a layer in which two homogeneous LLM agents share a single Docker container and Git tree. Each agent’s effects are auto-committed under its identity, and each next observation is prepended with a structured peer-activity block. The final submission is the team patch derived from the shared history. 

Results. We evaluated PASC on the full Python subset of SWE-Bench Pro using two independently-developed LLMs, against an isolated single-agent baseline and a "silent" two-agent baseline without the peer-activity feed. PASC delivers a statistically significant lift over the single-agent baseline on both models. Crucially, the silent baseline is statistically equivalent to the single-agent one, confirming that the gain stems from action-space coordination rather than parallelism. PASC also outperforms a deployable federated alternative in which agents share the peer-activity feed but work in isolated worktrees. Relative to the silent baseline, PASC reduces cost per resolved task by ∼20% and destructive concurrent edits by ∼47%. Preliminary observations indicate that beyond two agents, active interference increases several-fold, suggesting that larger populations will require supplementary coordination mechanisms. 

Conclusions. A shared-workspace configuration of homogeneous agents coordinated through action-space observation improves effectiveness over an isolated single-agent baseline, and does so more economically than a comparable multi-agent baseline without peer awareness.