Google just dropped Harness² as open-source and it's actually wild.
Instead of manually configuring your AI agent setup once (prompts, guardrails, memory, tools, subagents, etc.), Harness² lets the agent rewrite its own config on the fly. No retraining. No human babysitting.
How it works:
→ Agent proposes changes to its own harness for the current task
→ Tests those changes in real-time
→ Keeps what works, no ground truth needed
→ Carries improvements forward to new tasks
Benchmarks:
→ GLM-5 and Gemini 3.5 Flash saw +11.5 pts on WorkBuddy-Bench
→ +11.8 pts on LAB (legal tasks)
Works with OpenCode or Codex. Apache-2.0 license. 11.3K stars already.
Demo runs local, no API keys:
git clone && cd harness2
uv venv --python 3.12 .venv && source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2
This is the kind of infra that separates alpha builders from prompt merchants.
Instead of manually configuring your AI agent setup once (prompts, guardrails, memory, tools, subagents, etc.), Harness² lets the agent rewrite its own config on the fly. No retraining. No human babysitting.
How it works:
→ Agent proposes changes to its own harness for the current task
→ Tests those changes in real-time
→ Keeps what works, no ground truth needed
→ Carries improvements forward to new tasks
Benchmarks:
→ GLM-5 and Gemini 3.5 Flash saw +11.5 pts on WorkBuddy-Bench
→ +11.8 pts on LAB (legal tasks)
Works with OpenCode or Codex. Apache-2.0 license. 11.3K stars already.
Demo runs local, no API keys:
git clone && cd harness2
uv venv --python 3.12 .venv && source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2
This is the kind of infra that separates alpha builders from prompt merchants.