Google Cloud AI Research reports that per-task harness rewriting moved a weaker model past a stronger one on WorkBuddy-Bench

The method freezes the model and edits eight harness components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. For each task it runs the agent once, proposes two edits (one for general agent failures, one for the task's domain), runs both on the same task, and composes a successor harness from how the runs differed. No ground-truth verifier is involved, and the authors say these contrasts "inform" composition rather than verify correctness. After each task, a library of adopted harnesses and lessons carries forward to the next one.

On WorkBuddy-Bench, with the default editor Claude Opus 4.8, one improvement step took GLM-5 under OpenCode from 64.2 to 75.7, a gain of 11.5 points, and Gemini 3.5 Flash under OpenCode from 72.4 to 77.5. That puts GLM-5 above Gemini's stock score. Parallel recursion at k = 5 took Gemini to 80.8. Reported gains elsewhere were up to 4.5 points on JobBench and 11.8 percentage points on the LAB legal benchmark. The authors also report that Harness Scaling, a per-task baseline, scored below the stock harness in seven of twelve benchmark and configuration pairs. Ablations cost 1.6 to 3.1 points on WorkBuddy-Bench when the library, scoped editors or composition step were removed.

The paper's JobBench case study is the most transferable part. A Gemini 3.5 Flash run scored 43.1. The trajectory showed procedural failures: figures never reconciled across source files, and outputs never reopened after writing. The fix was a guardrail to inventory inputs and trace figures to their sources, a script that reopens produced files, and a reconciliation skill. The rerun scored 77.8. The authors also report that using each model as its own editor improved all five models they tried, from GLM-5 to Claude Opus 5, at k = 2.

These are the authors' own results, means over three runs, with standard deviations of several points on some cells (the Gemini one-step average is 77.5 with a reported 2.5). Larger k costs more compute, though the paper says it stays on the cost-performance frontier. The repository warns that executor network access is unrestricted and that public benchmark rubrics may be reachable online, so anyone reproducing the numbers should lock down networking first. The code's local demo needs no model credentials. We did not run the benchmarks and did not read the appendices.

Why it matters: The measured gain came from edits you can make by hand today, not from the framework. In the case study the base agent never reopened its own output, and a guardrail plus a verification script changed that. If your harness has no step that rechecks produced files against their sources, that is the first edit this paper supports, and it works on whichever model you already run.