Google Open-Sourced Harness²: Agents That Rewrite Their Own Scaffolding

Verdict: Google Research open-sourced Harness² (Apache-2.0, arXiv:2610.14224), a framework that lets agents recursively rewrite their own scaffolding — prompts, tools, memory, skills — while model weights stay frozen. Verified from the README: one Harness² step lifts GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 points on LAB's partial pass rate. My read: this is the systematization of a trade-off every API buyer already faces — pay for a bigger model, or pay tokens for a smarter loop around a cheap one. Harness² is Google betting publicly on the second option, days before its flagship successor launch.

What it actually does

A "harness" is everything around the model that makes it an agent: the prompts, the tools it can call, its memory layout, its skill library. Harness² optimizes that layer with two connected loops. Task-level recursion runs a contrastive Propose–Probe–Compose step: propose edits to the harness, execute probe variants on the same task, and keep what behaviorally worked — notably, without any ground-truth verifier, which is the honest constraint for open-world tasks where nobody grades the output. Harnessing-level recursion then distills what worked into a persistent "improvement harness" that shapes the next task's edits. The first loop buys you a better harness today; the second is the compounding part.

It's plug-and-play in the unglamorous sense that matters: it runs on OpenCode and Codex executors, against three professional benchmarks (JobBench, LAB, WorkBuddy-Bench), installable via uv pip install. Chinese coverage (量子位) adds that after wiring GLM-5 into it, office-benchmark scores climbed — consistent with the README's claims.

The RSI line suddenly has receipts

Two days ago we covered the November leak cycle, including Google engineer Vedant Misra replying "have you heard of recursive self-improvement?" when internal Carbon scores surfaced. Here is the context that reply was sitting on: Google had already shipped — open, reproducible, Apache-2.0 — a working RSI loop for the harness layer. The调侃 wasn't a vibe; it was product roadmap. Whether Carbon itself ships with harness-style self-tuning is unknowable from outside, but the research line is no longer speculative.

This also reframes what we wrote about harness engineering in September: hand-writing a good harness was the 2025-2026 skill. Frameworks like Harness² turn that skill into a budget line — you're converting engineering hours into inference spend, automatically.

The cost math nobody puts in the README headline

The README does say it, one line down: "Model-backed runs incur API costs; start with a small task selection." Do the multiplication: task-level recursion means the executor runs the task multiple times per step — initial execution, plus one probe per proposed edit, plus composed re-runs. With k parallel probes, a single refinement step is easily 3-5x the tokens of one plain execution, before the harnessing-level reflection pass. On Gemini 3.5 Flash pricing that's cheap; on a frontier-model executor it stops being cheap fast. Harness² economically pairs with mid-tier models — which is exactly the configuration Google benchmarked (GLM-5, Gemini 3.5 Flash), and I'd treat the absence of Opus-class executors in the results as a tell, not an oversight.

Hold the gains loosely

Same rule we applied to OpenAI's COBOL numbers yesterday: vendor-selected benchmarks flatter the vendor. Google Research benchmarked its own framework on benchmarks built by Google-adjacent teams, against base harnesses they also chose. +11.5 points is real as reported, but "as reported" is doing work in that sentence. The honest posture: the mechanism is sound and open enough to audit, the numbers are plausible, and the only benchmark that matters is yours — which is runnable, because it's Apache-2.0 and the install is two commands.

For a solo developer my take is concrete: try it on Flash-class executors where probe costs are pennies, keep recursion budgets (the --k flag) small until you've watched a bill, and don't bother pointing it at tasks you already trust your harness on — the framework earns its tokens on tasks the base harness fumbles, which is also the only setting where its design makes sense.

Dated caveat (2026-10-21): framework facts and benchmark gains from the official GitHub README and arXiv abstract, both retrieved today; the 量子位 office-benchmark relay and Misra reply context are relayed coverage. I have not run Harness² myself — the cost multiplication is arithmetic from the documented pipeline (probes + re-runs + reflection), not a lived invoice. Expect the repo to move; it's a fresh research release.

Sources: google-research/harness2 (README, retrieved 2026-10-21); arXiv:2610.14224; harness2.github.io; 量子位 / BAAI relay coverage.

Related: November Triple Launch Leaks (the RSI quote) · DeepSeek Harness Quickstart · GPT-6.1 Sol COBOL Migration