Google code lets AI agents improve themselves without changing the model

Google Research has released Harness², a framework for making AI agents better without touching the language model itself. Instead, it tunes the agent's "harness": the prompts, tools, memory and control logic that surround the model.

The code is open under the Apache 2.0 license as research code. According to Wavect's review, it is version 0.1.0, released October 16, 2026, and it is aimed at developers, not everyday users.

Propose, probe, compose

The problem the researchers want to solve is unfamiliar tasks, where there is no ground truth to check the answer against. Here the framework repeats three steps: Propose (suggest changes), Probe (test them on the task at hand) and Compose (combine the improvements based on differences in behavior between runs).

At a second level, the framework stores a library of lessons and well-functioning setups across tasks, which guides later improvements. Eight components can be edited, including the system prompt, memory, skills and plugins. The process can run in parallel for broader exploration or sequentially for deeper refinement.

Results with caveats

According to the research paper, results rose by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB, which covers legal work. The figures were measured with the models GLM-5 and Gemini 3.5 Flash and rose further when more compute was used.

The figures come from the authors' own tests in a paper that has not yet been peer reviewed. Wavect stresses that these are maximum improvements for particular combinations of model and harness, not something that applies across the board. Which baseline the reported points are measured against has not been examined in detail.

Costs more and requires caution

According to Wavect, a recursive run with k=5 requires about 20 runs of the agent itself plus additional work with the improvement model, so exploration costs more than execution. Model-based runs require Linux, Python 3.12+, Node.js 20+ and configured cloud access. There is, however, a demo that needs no credentials.

Security is something to watch. The default setup does not restrict the agent's network access, and Codex runs with its sandbox turned off. Public benchmark rubrics may also be available online, so network restrictions are recommended before evaluation. Wavect recommends disposable environments, limited credentials and externally enforced network restrictions before connecting the framework to enterprise systems.