A Google Research paper rewrites the execution harness around a frozen model through two levels of recursion. One step lifted GLM-5 to 75.7 on WorkBuddy-Bench - above Gemini 3.5 Flash's 72.4 base score. The real finding is bigger than the leaderboard screenshot.
A Google Research paper released this month, which I read on October 23, 2026, introduces Harness²: a framework that rewrites the execution harness around a frozen model - prompts, tools, memory, orchestration - through two levels of recursion, without any ground-truth verifier. One improvement step lifted frozen GLM-5 to 75.7 on WorkBuddy-Bench, above Gemini 3.5 Flash's 72.4 base-harness score on the same benchmark. The comparison being shared online is real but easy to misread.
The paper - "Harness²: Recursive Agent Harnessing for an Open World", arXiv 2610.14224, submitted October 9, 2026, by Ruiyao Xu, Yanfei Chen and seven co-authors - starts from a simple observation: an agent is a model plus a harness, and the harness is usually hand-designed once, for tasks nobody has seen yet. Harness² treats the harness itself as the thing to optimize. Task-level recursion runs a contrastive Propose-Probe-Compose step: propose edits to the current task's harness, probe their effects on that same task, and compose a refined harness from behavioral contrasts between executions - deliberately without ground-truth feedback, because open-world tasks have no verifier at test time. Harnessing-level recursion then distills what worked into a persistent "improvement harness" that carries lessons to future tasks. The model weights never change. arXiv abstract
The result spreading on social media this week is real, and it is the paper's own: across WorkBuddy-Bench, JobBench and LAB, a single Harness² step improves frozen GLM-5 and Gemini 3.5 Flash over their base harnesses (off-the-shelf OpenCode and Codex), with gains up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB's partial pass rate. On WorkBuddy-Bench specifically, one step lifts GLM-5 to 75.7 with OpenCode and 75.3 with Codex - in the paper's words, "surpassing the corresponding base Gemini scores of 72.4 and 74.5". Note what that comparison is: an improved GLM-5 against Gemini's unchanged starting harness, not GLM-5 beating Gemini as a model. Given the same step, Gemini 3.5 Flash itself climbs from 72.4 to 77.5, and to 80.8 with five rounds of parallel recursion. The finding is not that one model wins; it is that a large fraction of what we call model capability lives in the scaffolding and can be improved mechanically.
The release is a real codebase, not just a paper: the repository is public under the Apache License 2.0 (I read the LICENSE file), with install and benchmark guides for JobBench, LAB and WorkBuddy-Bench, support for OpenCode and Codex executors, and both parallel and sequential recursion modes. The reported figures are the authors' own means with standard deviations, which is normal for a first-party research release; independent replication is the usual next step before treating the effect sizes as settled.
Most public model comparison quietly assumes the harness is fixed and the model is the variable. This paper inverts the experiment and finds double-digit movement from the part we were holding constant. For anyone building agents, the practical reading is not "switch models" but "your harness is a versioned, testable artifact with real points on the table" - and the paper hands over one working method for tuning it. That is a more useful takeaway than the leaderboard screenshot.