Google Researchers Publish Harness^2 Recursive Agent Framework

Researchers including authors affiliated with Google posted a paper to arXiv on October 9, 2026, describing a framework called Harness^2 that aims to let an AI agent’s execution harness improve itself while the agent is working on open-world tasks it has never seen before. An execution harness is the scaffolding around a model: the prompts, the tool-call sequences and the retry logic that decide how a raw model turns into something that finishes a task. The paper’s authors argue that most prior attempts at recursive self-improvement for harnesses stall in the open world because there is no ground-truth verifier sitting beside a novel task to tell the system whether an edit helped.

Harness^2 answers that with two connected loops. A task-level loop runs a contrastive step the authors call Propose-Probe-Compose, which edits the current task’s harness without needing a labeled correct answer, and can be run again in parallel or in sequence to push further gains. A second, harnessing-level loop then pulls lessons out of those edits and folds them into a longer-running memory, so later tasks start from an already-improved harness rather than from scratch.

The team tested the framework on four model-harness pairings, including Gemini 3.5 Flash and GLM-5, across three benchmarks aimed at professional work. A single editing step raised scores over each model’s unedited harness on every configuration tested, and the authors report that running the task-level loop for a larger budget produces further gains beyond that first step.

The numbers

The paper reports gains of up to 11.5 points on a benchmark called WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on a benchmark called LAB, all measured as the improvement a single recursive editing step produced over each model’s base harness, before any additional recursion budget is spent. The paper frames these as a floor rather than a ceiling, since it states that further recursion raises scores again.

The work sits in the growing body of research treating the harness, not just the underlying model, as the thing worth tuning automatically, which matters for anyone weighing whether to license a bigger model or invest in better scaffolding around the one they already have.