The system eliminates the traditional main agent, substituting it with a shared context and a task queue where agents claim tasks asynchronously to reduce idle time and redundancy. On ProgramBench, the framework beat Claude Code's test pass rate by up to 19.9 points while maintaining a 120 minute budget. The team is open-sourcing the code and all trajectories, enabling users to run the collaborating agents directly within Codex and Claude Code.
Wednesday, October 21, 2026
Story
Stanford DeLM Beats Claude Code Accuracy by 17.5 Percentage Points
AI12 days ago
A new coordination system for AI agents allows parallel workers to solve software engineering problems without a central controller. Researchers at Stanford released an updated version of DeLM, a decentralized framework that outperforms Codex and Claude Code on benchmarks including Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench. Utilizing GPT-6-Astra and Claude Opus 5.5, DeLM improved accuracy by up to 17.5 percentage points over the strongest baseline and ran up to 2.49x faster than the systems it builds upon.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.