A new open-source framework called DeLM is shaking up how we think about multi-agent AI systems by ditching the traditional central orchestrator. Developed by Yuzhen Mao and a team including Jerry Gu, Aadi Chauhan, Qizheng Zhang, Hangoo Kang, and Azalia Mirhoseini, DeLM allows parallel agents to coordinate asynchronously through a shared context and task queue. This decentralized approach aims to solve tasks faster and more accurately by letting agents claim work, publish findings, and reuse each other's files without a main agent relaying every move. The core innovation lies in the 'decentralized coordination' model. Agents interact with an append-only shared context where they post progress, failed approaches, and completed subtasks. Peers read these entries and import shared files into their own independent workspaces. If a defect is found, it's reported as a FAIL entry and fixed by new entries rather than edits, preserving an audit trail of the reasoning process. This structure allows agents to choose complementary work by seeing who owns which task in the common queue.

Benchmark Performance and Latency

The numbers suggest this architecture delivers tangible benefits. On a 10-task subset of Terminal-Bench 4.0, DeLM with two agents averaged a 12.5 percentage point accuracy increase and a 1.78x speedup over base harnesses across two models. Specifically, using GPT-6-Astra, DeLM (n=2) achieved 85.00% accuracy compared to 71.67% for the single-agent Codex harness, while reducing latency from 26.91 minutes to 17.47 minutes. With Claude Opus 5.5, the gains were even more pronounced in speed, hitting a 2.02x speedup with DeLM (n=2). On DeepSWE v1.1, DeLM continued to outperform single-agent baselines. For GPT-6-Astra, accuracy jumped to 98.33% with n=2, up from 88.33% for standard Codex, with a 1.25x speedup. Claude Opus 5.5 saw similar improvements, reaching 88.33% accuracy with DeLM (n=2) versus 71.67% for the baseline, alongside a 1.50x speedup. The framework also showed strong results on SWE-bench Verified, where DeLM (n=4) achieved 66.08% accuracy, significantly higher than the 49.32% of Claude Code.

Integration and Accessibility

DeLM isn't just a research artifact; it's designed for integration into existing developer workflows. The team released a plugin, delm-agent-plugin, that allows users to run DeLM's collaborating agents directly inside Codex and Claude Code. The setup requires Python 3.13+, Git, and Linux Docker Engine with BuildKit. The project provides independent Docker images for both Codex and Claude Code harnesses, ensuring that building one does not require building the other. All 720 agent trajectories from Terminal-Bench 4.0 and DeepSWE v1.1 evaluations are also available on Hugging Face for transparency and further analysis.

Key Takeaways

  • DeLM uses a decentralized, peer-to-peer coordination model via a shared context and task queue, eliminating the single point of failure or bottleneck of a central orchestrator.
  • The framework demonstrates significant performance gains, including a 12.5 percentage point accuracy boost and 1.78x speedup on Terminal-Bench 4.0 with two agents.
  • DeLM integrates with existing tools via a plugin for Codex and Claude Code, making it accessible for current development workflows without requiring a complete stack overhaul.
  • The system maintains an append-only log of progress and failures, allowing agents to learn from each other's mistakes and avoid redundant work.

The Bottom Line

DeLM proves that ditching the central controller doesn't mean losing control; it actually unlocks higher efficiency by letting agents collaborate directly. If you're building multi-agent systems, stop bottlenecking your workers with a master agent and start sharing context.