← News

Agents & MCPSources: arXiv · 19 Nov 2026

EnvAgent automates building Docker environments for scientific code repositories

A research team describes an agentic framework that analyzes repositories, writes Dockerfiles, and repairs failures, reporting stronger results than baseline coding agents on a new benchmark.

In plain words

Running AI agents on scientific code requires a working software setup, which is usually built by hand and takes significant developer effort. EnvAgent is a system that creates these setups automatically: it inspects a repository, plans the needed configuration, writes the container definition, and fixes errors when builds fail. The authors report it produced working environments far more often than existing coding agents they compared against.

What was announced

The paper argues that preparing execution environments is a bottleneck for AI-driven scientific computing, since repositories carry incomplete requirements for operating systems, toolchains, and dependencies, and current workflows depend on developer-maintained setup scripts. EnvAgent pairs deterministic utilities for Docker execution and state management with coordinated agents that handle repository analysis, planning, Dockerfile synthesis, and iterative failure diagnosis and repair, according to the abstract.

To measure performance, the authors introduce EnvBench, a benchmark of 54 scientific artifacts and open-source repositories across seven languages, drawn from Chameleon Trovi and Multi-SWE-bench, with 2,108 weighted checks applied by a shared evaluator. In four matched language-model settings, the authors report EnvAgent built an average of 43.5 executable environments out of 54, compared with 23.5 for Codex and Claude Code, and scored 62.02 versus 24.19 on a composite of executability and rubric satisfaction. These figures come from the paper's own evaluation and have not been independently confirmed.

What it does not show

  • The evidence is limited to the paper's abstract, so methodological details and full results are not available for verification.
  • The reported performance figures are the authors' own evaluation and do not constitute independent confirmation.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.