Google Research lets the agent fix its own harness

A "harness" is everything wrapped around a model that turns it into an agent: the system prompt, guardrails, memory, tool guidance, skills, scripts, subagents and plugins. Usually a person writes it once and every task runs through the same setup. Harness², a project published on Google Research's GitHub under the Apache 2.0 license, lets the agent rewrite that setup for each new task while the model weights stay frozen.

Here is how it works, according to the README and the paper. For each task the system runs the current harness once, then proposes two edited versions: one aimed at how the agent works (planning, tool use, checking its own output) and one aimed at what the domain needs. It runs both on the same task, compares how the runs behaved, and composes a final harness from the edits the evidence supports. No answer key or grader is involved. After the task, it saves the adopted harness and the lessons that held up into a library that guides the next task.

The reported results: with GLM-5 and Gemini 3.5 Flash as the working models, a single round of edits improved scores by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on the partial pass rate of Harvey's Legal Agent Benchmark. More rounds pushed scores higher. In the paper, the default editor making those changes was Claude Opus 4.8, a separate and stronger model, though the authors also show gains when one model plays both roles.

Who it is for: teams running agents on varied professional work (office documents, legal drafting, research tasks) who spend time hand-tuning prompts and skills, and researchers studying self-improving agents.

The caveats are in the project's own docs. It runs on Linux with Python 3.12. The local demo needs no model credentials, but real benchmark runs default to Vertex AI on Google Cloud and "incur API costs," and every round means extra full runs of the task plus editor sessions. The README warns that executor network access is unrestricted and Codex runs with its sandbox disabled, so you set the network limits yourself. The paper's limitations section says the feedback signal can be thrown off by run-to-run variance and shared model biases, and its failure analysis shows cases where the selector picked a weaker output. And the README says plainly: "This is not an officially supported Google product."

MODEL has not run Harness². Everything above comes from the repository, the paper and the project site, read on October 18.