Skip to main content
Ctrl+K

TEL Research Group

  • Research
  • Publications
  • People
  • Tao's CV
  • Software
  • FermiLink@AI
  • Gallery
  • News
  • GitHub
  • X
  • Research
  • Publications
  • People
  • Tao's CV
  • Software
  • FermiLink@AI
  • Gallery
  • News
  • GitHub
  • X

Section Navigation

FermiLink@AI#

We are trying to learn how to use AI agents effectively in scientific research. This page collects two small open-source contributions along this direction. Both are experimental and under active development, and we welcome feedback.

FermiLink harness

FermiLink logo

FermiLink is an agent harness for multidomain scientific simulations. Given a simulation goal in a plain goal.md, FermiLink can robustly run multi-day, multi-step simulation experiments by interfacing with different coding agents.

$ pip install fermilink
$ fermilink agent codex        # or: claude, gemini, opencode
$ fermilink install meep       # 150+ package knowledge bases
$ fermilink loop goal.md       # long-running simulations
Documentation GitHub PyPI arXiv:2604.03460

codex-fermilink

codex-fermilink forks from OpenAI's open-source Codex CLI, which is optimized for long-duration scientific workflows.

It brings the FermiLink harness within codex, so we can run FermiLink workflows with the same user experience as the original Codex CLI.

$ npm install -g @tel-research/codex-fermilink
$ codex-fermilink      # then type /profile for using scientific workflows. 
Terminal recording of codex-fermilink: the /profile command opens a picker, the Scientific Simulations profile is selected, and the agent runs a small convergence test.
Local demo on a laptop: /profile switches to the scientific profiles.

previous

Software

next

Gallery

© Copyright 2024-2026, TEL Research Group.