DeLM: decentralized code agents with shared context, without a central coordinator

What happened and what to do

The Stanford article describes DeLM, a system for long-running code agents that replaces a central coordinator with two elements: a context shared among the agents and a task queue. According to the author, the approach reaches a 2.49× speedup and higher accuracy compared with the Claude Code and Codex baselines. The code and experiment trajectories are open.

For a company already using code agents on long tasks, the practical question is whether central coordination is a bottleneck. One path is to build, on its own infrastructure, a shared context store, a task queue with access control, and a log of every change the agents make. The team can then run a controlled evaluation on its real repositories, measuring time to completion, error rate and quality of the generated code, using the same tasks for each architecture. The author's numbers serve as a hypothesis to test, not as a guaranteed result in your context.

How the consultancy can help

Diagnosis of how long-running code agents are orchestrated today in your team and whether a decentralized model, with shared context and a task queue, fits your cases. Then, a scoped implementation of an evaluation harness on real repositories, with queue, context storage and monitoring on your infrastructure, followed by supported operation.

Next step

If your team uses code agents on long tasks and wants to evaluate this alternative to central orchestration, send a short description of your case: the repositories involved, the tools in use and the coordination problem you face. We will reply with a scoped proposal.