T
Terminal-Bench-Science
VerifiedEvaluates agents on 70 expert-curated real scientific research workflows, run and graded through the Harbor benchmark harness.
- Access
- CLI
- Auth
- None
- Skills
- 3
Skills
Benchmark Run
Runs an agent against 70 curated scientific research tasks in reproducible containers through the Harbor CLI.
Task Grading
Grades agent transcripts against per-task rubrics and re-grades stored runs without re-executing the agent.
Result Comparison
Compares scores across models and harnesses on the same task set to track capability on real research workflows.
Research & Knowledgeagent-benchmarkscientific-workflowsharbor-frameworkagent-evaluationstanfordreproducible-grading
Visit Agent terminal-bench-science
Evaluates agents on 70 expert-curated real scientific research workflows, run and graded through the Harbor benchmark harness.
fields
nameTerminal-Bench-Science
providerStanford University
urlhttps://github.com/harbor-framework/terminal-bench-science
categoriesresearch
accesscli
authnone
streamingfalse
pushfalse
verifiedtrue
officialfalse
tagsagent-benchmark, scientific-workflows, harbor-framework, agent-evaluation, stanford, reproducible-grading
skills
benchmark-runBenchmark RunRuns an agent against 70 curated scientific research ta…
task-gradingTask GradingGrades agent transcripts against per-task rubrics and r…
result-comparisonResult ComparisonCompares scores across models and harnesses on the same…