Terminal-Bench-Science
by Stanford University researchers with the Terminal-Bench / Harbor team and contributing domain scientists
What is Terminal-Bench-Science? Terminal-Bench-Science is an agents & tool use eval made by Stanford University researchers with the Terminal-Bench / Harbor team and contributing domain scientists. A science-focused sibling of Terminal-Bench: agents work through research workflows written by practising scientists (data analysis, simulation, signal processing, theorem proving and more) inside sandboxed terminals, and each task is graded pass/fail by its author's own verifier tests. Release 0.1 (27 Aug 2026) has 70 tasks across life, physical, earth, mathematical and engineering sciences; the announcement reports a top resolution rate of 30% on its official leaderboard.
Editor note
A science-focused sibling of Terminal-Bench: agents work through research workflows written by practising scientists (data analysis, simulation, signal processing, theorem proving and more) inside sandboxed terminals, and each task is graded pass/fail by its author's own verifier tests. Release 0.1 (27 Aug 2026) has 70 tasks across life, physical, earth, mathematical and engineering sciences; the announcement reports a top resolution rate of 30% on its official leaderboard.
- What it tests
- Agents doing real scientific research workflows end to end in a terminal.
- Original source
- www.terminal-bench-science.ai/
- Last updated
- v0.1.0 tagged 2026-08-26; repo last push 2026-10-26
- License
- Apache-2.0
Related agents & tool use evals
- GAIA — General assistant skills: browsing, file handling, multi-step reasoning.
- WebArena — Autonomous web navigation and task completion.
- OSWorld — Computer-use agents operating real operating systems and apps.
- Terminal-Bench — Agents working in the shell end to end.
Discussion
No one has discussed Terminal-Bench-Science yet. Have you run it? Share what it caught, what it missed or how you set it up.