Loading patterns…
⌨️
Terminal-Bench(TB)
Hard command-line tasks in isolated environments, each with a human-written solution and tests that decide whether the agent actually finished.
Complexity: highEvaluation and Monitoring
In 30 seconds
- What
- Poses hard command line tasks in isolated containers, each carrying a human-written solution and tests that decide afterwards whether the work was actually finished.
- When to use
- You are evaluating agents that do infrastructure, build or debugging work in a shell, and you need correctness decided rather than judged.
- Watch out
- The score names a model and a harness together: the same weights under a weaker scaffold give up exactly where a stronger one reads the error and recovers.
Ask the AI expert about this pattern
Opens the assistant with your question prefilled. You review it before sending.
Terminal-Bench: Overview
Hard command-line tasks in isolated environments, each with a human-written solution and tests that decide whether the agent actually finished.
- One container per task, so runs cannot contaminate each other
- Every task carries a human-written reference solution
- Verified by tests rather than by inspecting the transcript
- Drawn from real infrastructure, data and debugging work
- Deliberately hard: frontier systems remain well short of solving it
- Harness-agnostic, so the agent and the scaffold are measured together
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.
Also known as: CLI agent benchmark, Command line agent evaluation
References
The papers, specifications, and repositories this pattern is based on.