Google Research's Harness² tunes an agent's prompts and tools, not its weights

A team at Google Research has published Harness², a framework that improves an AI agent by editing the system around the model rather than the model itself. The paper was submitted to arXiv on 9 October 2026, and the code is on GitHub at google-research/harness2 under Apache-2.0. This is a catch-up: the release is ten days old.

It is research code. The README ends: “This is not an officially supported Google product.”

What it changes

The README defines a harness as “the system around a model that enables it to act as an agent: prompts, tools, memory, skills, orchestration, and more.” (We explain the idea in What an agent harness is.) Harness² exposes eight editable parts: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Model weights stay fixed.

It works at two levels:

• Per task. A “Propose, Probe, Compose” step proposes edits to the harness, probes their effect by running the same task, and composes a refined harness from the differences between those runs, “without ground-truth feedback.” That last part is the point: real tasks rarely come with an answer key to score edits against.
• Across tasks. After each task, an “improvement harness” keeps the adopted harness and the lessons the evidence supports, to guide refinement on later tasks.

The per-task step can scale wide or deep. In parallel mode it probes several candidate harnesses at once. In sequential mode it carries the refined harness through successive steps. The --k flag sets that budget.

The reported results

According to the abstract, across three professional benchmarks and four model and harness setups, “a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses,” with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB’s partial pass rate. The authors say a larger recursion budget brings further gains.

These are the authors’ numbers, and they are maxima (“up to”). GLM-5 is Zhipu AI’s model. We have not reproduced any of it.

Trying it

The README asks for Linux, Python 3.12 and uv. After an editable install, harness2-demo --output runs/demo --k 2 runs both recursion modes “with deterministic editors and real Python executions” and, per the README, “requires no model credentials.” That is the cheap way to see how the pieces fit.

Model-backed runs are heavier. They need Node.js 20 or newer and the Google Cloud CLI, are configured for Vertex AI model IDs, and the README warns they “incur API costs”. The executor defaults to OpenCode, with Codex available through --substrate cx.

One caution from the README itself: “Executor network access is unrestricted, and Codex runs with its sandbox disabled.” It also notes that public benchmark rubrics may be online, and asks you to set network restrictions before evaluating so an agent cannot look up the answers. Run it in a disposable environment.

Why it matters

Most teams cannot retrain the model they use, but they do control the prompt, the tools and the memory around it. Harness² treats those as the thing to optimize, and tries to do it without a grader for each new task. Whether its edits generalize beyond these three benchmarks is the open question. For a picture of what a harness does turn by turn, see A day inside a harness.