Benchmarks and learning resources

Terminal-Bench

Terminal-Bench is a continuous benchmark of agent-model pairs on containerized terminal tasks with programmatic outcome verification. Harbor runs repeated trials and publishes versioned leaderboards.

Maintainer
Harbor Framework and Laude Institute
Deployment
Open benchmark run through Harbor
Status
Current as of September 22, 2026

What it does

Terminal-Bench is a continuous benchmark of agent-model pairs on containerized terminal tasks with programmatic outcome verification. Harbor runs repeated trials and publishes versioned leaderboards.

  • Terminal tasks spanning coding, systems, networking, security, dependencies, science, and interactive tools
  • Containerized environments, programmatic verifiers, repeated trials, and parallel execution
  • Versioned releases, public leaderboards, oracle validation, and task-quality review

Where it fits

Comparing agents on difficult, verifiable coding, systems, security, and scientific work inside sandboxed terminals.

Use it after defining the behavior that matters, collecting representative cases, and deciding how each case will be graded. A tool can execute and organize an eval, but it cannot decide which failures are costly for your users.

Limitations and cautions

  • Results are comparable only when the exact release, harness, agent, model, resources, and trial count are reported.
  • Terminal-Bench 2.1 corrected 28 of 89 version-2.0 tasks, showing sensitivity to dependencies, resources, networking, and sandboxes.

How to evaluate the evaluator

Before using any score as a release gate, create a small evaluator-validation set that humans have labeled. Compare false passes, false failures, disagreement by failure type, run-to-run variance, latency, and cost. Keep deterministic checks for verifiable facts, calibrate model judges against domain experts, and review raw traces when aggregate scores move.

Primary sources