Optimizes an agent's harness (prompts, tools, memory, skills, subagents) on unseen tasks through recursive Propose–Probe–Compose refinement, for researchers benchmarking agents.
Use it now
Keep it with up to 9 other tools you're setting up, then share them all as one link.
Try this first
Ask your coding agent:
- Task-Level Recursion
Use Harness² to propose harness edits: ‹what you need›
- Harnessing-Level Recursion
Use Harness² to update a persistent improvement harness: ‹what you need›
- Parallel and Sequential Recursion
Use Harness² to explore K candidate harnesses in parallel: ‹what you're looking for›
Install the CLI
Install it as its own instructions say.
Setup guide: github.com/google-research/harness2
Compare
Press Compare to add it to the tray at the bottom of the window, then pick up to 2 others anywhere on the site.
- Access
- CLI
- Auth
- None
- Skills
- 4
Skills
Task-Level Recursion
Proposes harness edits, probes their effects on the same task, and composes a refined harness without ground-truth feedback.
Harnessing-Level Recursion
Updates a persistent improvement harness with adopted harnesses and lessons from each task to guide later refinement.
Parallel and Sequential Recursion
Explores K candidate harnesses in parallel or refines one harness through K successive steps.
Benchmark Runs
Sets up, runs, judges, and reports JobBench, LAB, and WorkBuddy-Bench tasks with OpenCode or Codex executors.