Harness2: recursive agent harness
Harness² refines prompts, tools, memory, and skills around a fixed model. Apache 2.0. Not an officially supported Google product.
Key features
- Task-level recursion proposes edits, probes them on the same task, and composes a harness
- Harnessing-level recursion keeps adopted harnesses and lessons for the next task
- Eight editable parts: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins
- Local demo uses Linux, Python 3.12, and uv, and needs no model credentials
- Model-backed benchmarks use Vertex AI, with OpenCode and Codex executors
- Ships JobBench, LAB, and WorkBuddy-Bench
- README reports gains for GLM-5 and Gemini 3.5 Flash on those three benchmarks
- Apache 2.0. The README says this is not an officially supported Google product
- Category
- Agents
- Who it's for
- Engineers refining an agent harness for tasks unseen when that harness was designed.
- Tags
- agentharnesslocal-inferencepython
- Website
- github.com
- Listed
- 2026-10-26
Frequently asked questions
Is Harness² free?
Yes. Harness² is open source and free to use.
What is Harness² used for?
Framework that refines an agent harness on new tasks while model weights stay fixed. Apache 2.0. Typical uses: task-level recursion proposes edits, probes them on the same task, and composes a harness; harnessing-level recursion keeps adopted harnesses and lessons for the next task; eight editable parts: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.
What are the best Harness² alternatives?
Alternatives in the same category include Agensh, LearnHarness, OpenShell, Haystack, Alter.
Related in Agents
See all →- Agensh Self-organized multi-agent harness. Agents claim sub-tasks and merge progress without a central orchestrator. MIT.
- Alter Privileged access for AI agents. Vaulted credentials, per-call policy, and an audit log. Free tier is 10,000 API calls a month.
- LearnHarness Agent harness that plans and teaches from your current understanding, then keeps a knowledge graph. AGPL-3.0.
- Omnara Open-source agent API for any model and machine. Self-host free under Apache 2.0, or use the cloud with a $0 platform fee.
- OpenShell Safe, private runtime for fleets of autonomous AI agents. Sandboxes enforce file, syscall, and network policy. Apache 2.0.
- Agentic Batch Changes Describe a change in plain language. The agent scopes it across your indexed codebase and tracks it through to merge.