Decentralized Multi-Agent Coordination and Communication in Shared Codebases
| Source: arXiv AI
Tags: ParallACT, Claude Code, Claude Opus 4.8, Multi-Agent Coding, Coding Agents, Benchmarks
ParallACT, a new benchmark for coding agents working independently in the same repo, shows success on contradictory tasks dropping from 62% to 16% for Claude Code with Claude Opus 4.8, and communication channels recover little.
Details
Coding agents in shared repositories often get separate requests from different developers. Their changes can each work alone and still conflict when merged. ParallACT contains complementary tasks, where agents must share an interface, and contradictory pairs, where both must edit the same code in incompatible ways. The authors compare a single-agent oracle handling all tasks with parallel agents under different visibility, ordering, and communication settings. Parallel agents did substantially worse, most sharply on contradictory tasks, where Claude Code with Claude Opus 4.8 fell from 62% to 16%. The same gap showed up across six further harness and open-weight model pairings. Telling an agent a peer exists did not help much, and a communication channel recovered only a small part of the loss at extra token cost.