Google Harness²: Recursive Agent Harness Without Weight Updates
Deep dive into google-research/harness2—Propose-Probe-Compose loops that improve agent harnesses without fine-tuning model weights, plus how teams can run the open-source cookbook safely.
Posted by
Related reading
Composio Instant: The Agent Tool Wallet, Revisited After Its Product Hunt Resurgence
Composio Instant gives agents a prepaid wallet and 1,500+ MCP tools—we revisit the Product Hunt resurgence, wallet-vs-virtual-card trade-offs, and spending guardrails.
Eleven Realtime vs OpenAI Realtime API: Inside ElevenLabs' All-in-One Speech-to-Speech Bet
ElevenLabs’ Eleven Realtime pitches one speech-to-speech model on an OpenAI-compatible API—what that means for latency, voice quality, and vendor switching.
T3 Code: The Open-Source Coding Agent Control Plane, and What Its Latest Update Changes
T3 Code turns scattered coding agents into one control plane—what the late-October update adds for remote iOS simulators, connect speed, and shared threads.
Why harness engineering broke out in 2026
Teams shipping coding and ops agents hit the same wall: the base model is “good enough,” but tool routing, memory, and guardrails are not. Tweaking prompts by hand does not scale, yet fine-tuning for every workflow is expensive and brittle. Google Research’s Harness² (Harness2) project argues for a third path—recursively improve the harness around a frozen model.
The open repository google-research/harness2 and site harness2.github.io document a Propose → Probe → Compose loop: propose harness edits, probe them with targeted evaluations, then compose the winning changes into the next harness generation. The associated paper is arXiv 2610.14224.
Propose-Probe-Compose in plain language
- Propose: a meta-agent drafts candidate changes—new tool policies, memory schemas, or sub-agent splits.
- Probe: cheap, task-specific probes score each candidate without deploying to production traffic.
- Compose: merge winners into the live harness config and repeat.
Because weights stay fixed, the loop plays nicely with vendor models you cannot fine-tune anyway. The tradeoff is compute spent on probes and the quality of your evaluation harness—garbage probes still produce confident bad harnesses.
Practical adoption checklist
Before you run recursive harness jobs on production repos, pair research tooling with shipping guardrails already cataloged on AIXList: Sphinx Understand for human comprehension gates on agent-written code, agent tool-call guards for policy enforcement, and Beacon when you want durable skills extracted from long agent sessions.
For teams comparing hosted eval products, read our agent context workspace hub—shared context is the other half of the harness problem.
FAQ
Is Harness2 production-ready?
Treat it as research software: pin commits, reproduce paper benchmarks, and add your own probes for domain tasks before trusting autonomous harness edits.
Will Harness2 replace prompt engineers?
It automates search over harness configurations, but humans still define probes, safety constraints, and release criteria—especially for regulated workflows.