Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableKDnuggets→ original

Google Releases Harness²: An Agent That Rewrites Its Own Prompts and Scripts, Up to +11.8 pp on LAB Criteria

Google Research has released Harness², a framework in which an agent rewrites its own "harness" around the model: prompts, tools, scripts, while leaving the weights untouched. Like an intern who revises their instructions after mistakes, the agent runs a task with different edits and keeps the ones that help. On the LAB legal test, the share of passed criteria rose by up to 11.8 pp (say, from 50% to 61.8%). For teams using API-based models, this is a way to improve agents without fine-tuning, but the numbers so far come only from Google.

AI-processed from KDnuggets; edited by Hamidun News
Google Releases Harness²: An Agent That Rewrites Its Own Prompts and Scripts, Up to +11.8 pp on LAB Criteria
Source: KDnuggets. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceKDnuggets ↗
Editor-in-Chief
Key metric #1One step costs 4k agent runs and 3k improver sessions: at k=3, that is 12 and 9
Key metric #2In a demo without an API key, the agent fixed a bug in a script: the sum went from 105.25 to 85.15
All changes are plain filesthey can be read and checked against the original

Imagine an agent that processes accounting files, while a helper script in it quietly contains a bug: it skips negative amounts. Usually an engineer catches this: they look at the wrong total, fix the script or prompt, and run it again. Google Research has released Harness² as open source, a framework that automates this loop. It rewrites not the model itself but everything around it: instructions, tools, memory, and scripts. Google measured results in "percentage points" (pp): if an agent passed 50% of verification criteria and then passed 61.8%, that is a gain of 11.8 pp. That maximum was achieved on the LAB legal test.

What the "harness" is and what Harness² changes in it

A language model on its own is not yet an agent. What makes it an agent is the "harness": system instructions, rules on what is allowed and forbidden, long-term notes, skill sets for typical tasks, descriptions of how to use tools, helper scripts, subagents, and intermediate plugins. Today, most teams configure all of this by hand. For tasks you anticipated in advance, the approach works. But when the agent runs into something unforeseen, manual configuration cannot keep up.

Harness² automates this configuration while the model's weights stay frozen. There is no fine-tuning and no new training data. That makes the approach suitable for teams that work with the model only through an API: they need neither GPUs nor access to the weights.

Two loops: how the agent learns from its own runs

The "squared" in the name refers to two levels of recursion. The first loop works inside a single task using a propose–test–compile scheme. A harness editor proposes changes: general ones and ones specific to the domain. The agent solves the same task again with each version of the harness. Then a "composer" compares the runs and merges the edits that helped. No reference answer is needed, only the difference between two runs of the same task. In real work, nobody gives the agent a key with correct answers, so this is an important property.

The second loop works across tasks. The accepted harness and confirmed "lessons" go into a persistent experience library, and the next task starts with these notes. The first loop improves the agent for the current task; the second improves the improvement process itself.

The budget is set by the parameter k. In parallel mode, the system proposes k general and k specific edits and tries all of them: it searches wider. In sequential mode, it takes k steps in a row and carries the refined harness forward: it digs deeper. Each task costs 4k agent runs and 3k improver sessions, so at k=3 that is 12 runs and 9 sessions. LAB alone has 1,143 tasks, so a full run will not be cheap.

Numbers and the demo

Google tested GLM-5 and Gemini 3.5 Flash in four model–harness combinations across three benchmarks. A single improvement step raised the results of both models relative to the original harness. The largest gain on LAB is up to 11.8 pp in the share of passed evaluation criteria. These are best-case results, not averages, and they come from Google itself, so independent verification is still pending. Tables for each configuration are in the repository.

The demo runs locally in a couple of seconds and requires no keys. You need the uv package manager and Python 3.12. The original script sums the entries of an accounting ledger but filters out the negative ones. On a test file of three entries, it outputs 105.25. After Harness² works on it, the filter is removed, and the script computes the correct 85.15 (100.25 − 20.10 + 5.00). Nobody told the system the correct sum. A note remained in the experience file: when summing credits and debits, both signs must be taken into account, with high confidence. That entry carried over into the demo's second task.

All changes are stored in ordinary files: the rewritten prompt, the corrected script, the new rule. They can be read, compared against the original version, and accepted through code review. The framework even records why it accepted each edit.

A real run is more complicated and costs money. It requires a Google Cloud project with Vertex AI, as well as Pandoc and Poppler for the LAB documents. The .env configuration file has to be created manually: at the time of publication, it is not in the repository. The executor, improver, and judge are configured with separate model IDs. The author of the review advises starting with a single task so that paid calls do not pile up.

The practical takeaway: if you have an agent on someone else's model and manual prompt tuning has become tedious, Harness² offers a reproducible way to automate it. Start with the free demo, then compare the original and improved harness on one of your own tasks and judge whether the cost of the runs pays off.

Story

Key Questions Answered

What the "harness" is and what Harness² changes in it?

A language model on its own is not yet an agent. What makes it an agent is the "harness": system instructions, rules on what is allowed and forbidden, long-term notes, skill sets for typical tasks, descriptions of how to use tools, helper scripts, subagents, and intermediate plugins. Today, most teams configure all of this by hand. For tasks you anticipated in advance, the approach works. But when the agent runs into something unforeseen, manual configuration cannot keep up.

What are the key details on Two loops: how the agent learns from its own runs?

The "squared" in the name refers to two levels of recursion. The first loop works inside a single task using a propose–test–compile scheme. A harness editor proposes changes: general ones and ones specific to the domain. The agent solves the same task again with each version of the harness. Then a "composer" compares the runs and merges the edits that helped. No reference answer is needed, only the difference between two runs of the same task.

What are the key details on Numbers and the demo?

Google tested GLM-5 and Gemini 3.5 Flash in four model–harness combinations across three benchmarks. A single improvement step raised the results of both models relative to the original harness. The largest gain on LAB is up to 11.8 pp in the share of passed evaluation criteria. These are best-case results, not averages, and they come from Google itself, so independent verification is still pending. Tables for each configuration are in the repository.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…