TL;DR — Key Takeaways
- Better AI agents without model retraining. Google’s Harness² optimizes the prompts, tools, memory, skills and orchestration logic surrounding an AI model while leaving its weights unchanged.
- Early benchmarks show measurable improvements. Google’s research reports gains of up to 11.8 percentage points across professional benchmarks, although independent replication is still needed.
- Self-improvement creates new governance challenges. Automated harness changes introduce additional API costs, security considerations and the need for human review, particularly when agents can modify their own guardrails.
Google Research has released Harness², an open-source framework that improves AI agents by rewriting the scaffolding around a model instead of the model itself. The code is on GitHub under an Apache 2.0 license.
The idea is simple. An agent’s performance depends on more than the model. It also depends on the harness: the prompts, tools, memory, skills, and orchestration logic that turn a model into something that can do work. Harness² refines that harness while the agent runs. Model weights never change.
That matters for teams that can’t fine-tune, or don’t want to. Most enterprises use models through an API. They can’t touch the weights. Harness² lets them tune what they can touch.
How It Works
Harness² exposes eight harness components as editable files: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. Together, they control how context is built, how the agent calls tools, how it delegates work, and how it behaves inside its tool loop.
The framework runs two levels of recursion.
The first is task-level. Google calls it Propose-Probe-Compose. The agent runs a task once with its base harness. An editor model then proposes two candidate edits. One targets general agent failures, such as poor planning or tool misuse. The other targets domain knowledge and procedures for the task at hand. Both candidates run against the same task. A composer compares the results and keeps only the edits backed by observed behavior.
The second is harnessing-level. After each task, a reflector stores the adopted harness and a lessons-learned file. That library shapes future proposals, so the system improves over time. Each new task still starts from the base harness.
Here’s the key design choice. Most self-improvement methods assume you have labels, test cases, or a verifier that says whether an answer is right. Real work rarely comes with one. Harness² is built for tasks the agent hasn’t seen before. It uses the contrast between runs instead of ground truth.
Teams can scale the search with a budget setting. Parallel mode explores several candidate harnesses at once. Sequential mode chains refinement steps to go deeper.
What the Numbers Show
The research team tested Harness² on three professional benchmarks: WorkBuddy-Bench (office, web and code tasks), JobBench (professional work) and Harvey’s LAB (legal work). Executors were Gemini 3.5 Flash and GLM-5, running under the OpenCode and Codex agent harnesses. Claude Opus 4.8 served as the default editor model.
A single refinement step improved all four model-harness setups on all three benchmarks. The biggest gain came from GLM-5 under OpenCode, which rose from 64.2 to 75.7 on WorkBuddy-Bench. That’s 11.5 points. On LAB, gains reached 11.8 percentage points. The smallest gain was 1.2 points, for GLM-5 under Codex on JobBench.
Weaker models gained the most. After one step, GLM-5 outscored the base Gemini 3.5 Flash setup on WorkBuddy-Bench. And on that benchmark, three of four setups also finished tasks in fewer steps.
These are first-party results from the paper’s authors. No one has independently replicated them yet. The paper also compares Harness² with other harness-optimization methods, including Stanford’s Meta-Harness. But the Harness² team ran those comparisons, not the original authors.
“If these results survive independent replication, a harness edit carries the same production risk as a model upgrade and needs the same change review,” said Mitch Ashley, vice president and practice lead for CIO & Technology Buyers and Software Lifecycle Engineering at The Futurum Group. “The harness is the part most enterprises control.”
The Trade-Offs
Harness² isn’t free. Every refinement step means extra runs of the same task: the base run plus a probe for each candidate. Raise the budget and the run count climbs. Google’s README warns that model-backed runs incur API costs and suggests starting with a small set of tasks.
There are security notes, too. The README states that executor network access is unrestricted by default and that Codex runs with its sandbox disabled. Teams need to set their own network restrictions. That’s acceptable for a research tool. It’s not acceptable in production.
It’s also version 0.1.0, and Google labels it as an unofficial Google product. Setup requires Linux, Python 3.12, and the uv package manager. Model-backed runs go through Vertex AI. A local demo runs without model credentials, so it’s easy to try.
Why It Matters
Harness engineering has become one of the most time-consuming parts of building agents. Teams spend weeks tuning prompts, writing tool descriptions, and building skills files by hand. Harness² suggests much of that work can be automated. And in some cases, the payoff rivals a model upgrade.
But an agent that rewrites its own guardrails raises a governance question. If the harness changes per task, someone has to review what changed and why. Harness² saves adopted harnesses, reflection notes, and lessons as files. That’s a start, but files alone don’t make a review process.
“An agent that rewrites its own guardrails shifts the developer’s job from hand-tuning the harness to setting rules for what the agent may change,” Ashley said. “Saved files become an audit trail only when a named owner reviews them. Watch whether approval gates for harness edits arrive before this reaches production.”
For now, Harness² is a research release, not a product. But the direction is clear. The next gains in agent performance may come less from bigger models and more from better systems around them.
