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

Multi-agent Collaboration with State Management

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 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.

Proposes STORM, a multi-agent framework where engineers share one workspace and a state layer validates each write against file versions the agent has read (write-time conflict control plus intent annotations), outperforming git-worktree isolation on Commit0-Lite and PaperBench Code-Dev.

Problem

Concurrent LLM agents editing a shared codebase can silently conflict; workspace isolation with git worktrees defers conflict resolution to a costly post-hoc merge.

Approach

Replace workspace isolation with local state consistency: all agents share one workspace, and a STORM layer mediates file reads/writes, accepting a write only if the target and its read dependencies are unchanged since the agent last observed them, otherwise rejecting it with the current content, diff, and stale dependencies; intent comments coordinate semantics at shared boundaries.

Key results

• On Commit0-Lite with Sonnet 4.6, STORM reaches 46.2 weighted / 82.5 macro score vs. 24.6 / 63.8 for GitWorktree and 20.7 / 66.4 for single-agent.
• On PaperBench Code-Dev, STORM scores 74.1 with Sonnet 4.6 vs. 72.7 for GitWorktree and 68.7 for single-agent.
• STORM-Combined reaches 87.6 on Commit0-Lite and 78.2 on PaperBench with Sonnet 4.6.

Surprising findings

• Soft isolation (prompt-level constraints) performs about as well as git worktree isolation on Commit0-Lite weighted score (24.0 vs. 24.6), suggesting the isolation mechanism matters less than the choice of delegation.
• On PaperBench, STORM is not uniformly best: GitWorktree wins on several papers and single-agent on others; Qwen and DeepSeek single-agent results are competitive or better than GitWorktree on some metrics.

Limitations

• Gains are measured with proxy coupling and conflict metrics derived from manager-review events rather than manually adjudicated root causes.
• STORM cannot guarantee that task boundaries are semantically correct or that independently accepted edits compose cleanly; many failures occur after writes are accepted.
• PaperBench evaluated only on Code-Dev subset with an LLM judge and 20 papers, limiting statistical power; single-agent is often cheaper.

Experimental setup

Datasets: Commit0-Lite PaperBench Code-Dev

Benchmarks: Commit0-Lite PaperBench Code-Dev

Metrics: Score_w (weighted pass rate) Score (macro pass rate) Cost_eff Time_eff PaperBench judge score

Baselines: Single-agent GitWorktree GitWorktree-Combined Soft Isolation

Builds on: OpenHands optimistic concurrency control

Closest papers by meaning

• AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators (NeurIPS 2026 · similarity 0.85)
• When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding (NeurIPS 2026 · similarity 0.85)
• SyncMind: Measuring Agent Out-of-Sync Recovery in Collaborative Software Engineering (ICML 2025 · similarity 0.85)
• MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution (NeurIPS 2024 · similarity 0.84)
• DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.84)
• Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.84)