Skip to main content
​

​
New Harbor docs

Harbor docs now cover previously undocumented features such as simulated users, streaming, and regrading, and expose an MCP for coding agents. Read more.
​

​
Watch agents as they run

Harbor streams the agent’s trajectory and sandbox files while a trial is in progress. Read more.
​

​
Terminal-Bench 4.0

Terminal-Bench 4.0 calibrates task resources, applies task fixes, and removes saturated tasks on Harbor Hub. Read more.
​

​
Terminal-Bench-Science

A Stanford-led Harbor benchmark for scientific research workflows: 70 tasks across life, physical, mathematical, earth, and engineering sciences, with Opus 5 at about 30%. Read more.
​

​
Simulated users

A user agent can role-play a human and drive the target agent through a multi-turn conversation over ACP. Read more.
​

​
Regrade without rerunning the agent

harbor regrade runs a new grader on recorded trials so you can iterate on verifiers without paying for another agent session. Read more.
​

​
Load a previous trajectory into a new eval

--load-trajectory seeds a Harbor run with a prior Claude Code, Codex, or ATIF session so context moves into the next trial. Read more.
​

​
Interview an agent after the eval

harbor trial handoff pulls a finished trial into local Claude Code so you can inspect the work and ask follow-up questions. Read more.
​

​
Terminal-Bench 3.0 (formerly Frontier-Bench)

A living 74-task benchmark from the Harbor team where the best agents scored about 34%, runnable from the Harbor CLI and Hub. Read more.

​
Continuous Benchmarks

Benchmarks are not static artifacts. They are software and we should maintain them like software. Read more.
​

​
Hosted leaderboards on Harbor Hub

Publish dataset leaderboards on Harbor Hub, including reward-hacking detection on Terminal-Bench 2.1. Read more.
​

​
Harbor-Index

Compose the hardest Harbor tasks into an 82-task meta-benchmark that no agent cleared above 30% at launch. Read more.
​

​
Introducing harbor exec

Run an agentic map-reduce over sandboxes to analyze traces, mine sessions, and search across completed jobs. Read more.
​

​
Sidecar environments

Collect artifacts from sidecars, run verification in a sidecar, and execute pre-verification commands in any service. Read more.
​

​
Evaluate ACP agents natively

Harbor can run ACP registry agents with harbor run --agent acp:junie, adding 36 agents now and every future ACP agent when it lands in the registry. Read more.
​

​
Stop zipping your job results

Upload and share Harbor job results on Harbor Hub instead of zipping and sending them manually. Read more.
​

​
Verify Harbor tasks in a separate sandbox

Run task verification in a sandbox separate from the agent, with explicit artifact handoff between the two environments. Read more.
​

​
Multi-step tasks

The first major addition to the Harbor task format: tasks split into sequential steps, each with its own instruction and verifier. Read more.
​

​
Explore the Harbor cookbook

Recipes for building Harbor tasks and optimization loops. Read more.

​
The Harbor registry is getting an upgrade

A new way to package and distribute Harbor tasks and datasets. Read more.