Harness²
Research framework that improves an agent harness for each task.
- AI tool
- Free
Pricing
Free and open source. License: Apache-2.0.
Overview
Harness² is a research framework from Google Cloud AI Research and Northwestern University for improving the setup around a model, meaning its prompts, tools, memory and orchestration, rather than the model weights. For each new task it suggests edits, tests them on that task and combines what worked into a refined setup, and it carries what it learns forward to later tasks. The release can use Codex or OpenCode to execute tasks and ships guides for three benchmarks: WorkBuddy-Bench, JobBench and LAB. The accompanying paper reports that one improvement step raised scores for both models it tested.
Key features
- Task-level propose, probe and compose step
- Persistent improvement harness across tasks
- Parallel or sequential recursion
- Codex or OpenCode as the executor
Use cases
Alternatives
Last updated 2026-10-20