Benchmarks

← All benchmarks

Snorkel Agentic Coding 2.0

Code

Pass@1 on Snorkel's 200 frontier terminal tasks across nine task types and nine languages. Harbor evaluates deterministic tests, required outputs and trajectory rubrics in network-disabled Docker environments, with 30-minute agent and verifier limits and five attempts per task. Native model effort is not disclosed.

Top models (higher is better)