Research map / Multi-agent systems and games / LLM agents

When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Formalizes multi-agent coding orchestration as a communication-vs-critical-path graph partitioning problem and proposes Co-Coder, which partitions repository dependency graphs via hub isolation, Infomap, and latent-parallelism lifting, executed by a dependency-aware scheduler.

Problem

Adding parallel LLM coding agents introduces inter-agent communication overhead that can offset speedups and hurt code quality on repository-level software tasks with cross-file dependencies.

Approach

Treat multi-agent orchestration as partitioning a weighted file dependency graph to minimize critical-path computation plus cut-edge communication cost, using hub isolation, Infomap community detection, latent parallelism lifting, and a shared-task-list scheduler without global barriers.

Key results

• On DevEval, Co-Coder raises average pass rate from 56.8% (sequential) to 68.1% with 0.55x normalized latency and 0.72x normalized cost.
• On CodeProjectEval, Co-Coder reaches 34.1% average pass rate vs 20.1% sequential, with up to 2.10x wall-clock speedup and cost reduced by 35%.
• Claude Code w/ Agent Teams is fastest on CodeProjectEval (680 s) but has the lowest pass rate (16.3%), below sequential.

Surprising findings

• Claude Code with Agent Teams achieves lower pass rate than the sequential baseline on CodeProjectEval despite being fastest.
• File-based parallel gives negligible pass-rate gains on DevEval while inflating cost by 44%.

Limitations

• Near-complete coupling collapses the partition into one group, degrading to sequential execution.
• Evaluation limited to Python repositories.
• Hub and Infomap hyperparameters are heuristic and the cost model is approximated by static features.

Experimental setup

Datasets: DevEval CodeProjectEval

Benchmarks: DevEval CodeProjectEval

Metrics: pass rate wall-clock latency API cost

Baselines: Sequential coding agent File-based Parallel Claude Code w/ Agent Teams

Builds on: Infomap OpenHands Claude Code

Closest papers by meaning

• Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs (NeurIPS 2026 · similarity 0.90)
• AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.89)
• EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · similarity 0.88)
• A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems (ICML 2026 · similarity 0.88)
• Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows (NeurIPS 2026 · similarity 0.88)
• KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems (NeurIPS 2025 · similarity 0.88)