Google Releases Harness²: AI Agent Tunes Its Own Harness, GLM-5 Score Rises from 64.2 to 75.7
Google Research has released Harness², a framework in which an AI agent edits its own "harness": instructions, skills, and rules for working with tools. For example, if the agent greedily reads an entire repository, the editor adds a rule to look at the folder structure first. On WorkBuddy-Bench (office, web, and coding tasks), GLM-5's score rose from 64.2 to 75.7 points without fine-tuning. So before switching to an expensive model, it's worth checking what the harness configuration can deliver.
AI-processed from Habr AI; edited by Hamidun News
Google Research has published Harness², a framework in which an AI agent configures its own "harness" by itself. The harness is everything surrounding a language model: system instructions, rules for working with tools, memory, skills, and subagents. The agent equals the LLM plus the harness, and the harness determines whether the agent carefully searches for the files it needs or reads the entire repository and fills up its context. Engineers usually tune these settings by hand, which is slow and expensive. Harness² does this automatically. On WorkBuddy-Bench (office, web, and coding tasks), the score of GLM-5 paired with OpenCode rose from 64.2 to 75.7 points, a gain of 11.5. The model weights do not change.
What the Harness Consists Of
The framework's README lists eight components, and Harness² can edit each of them:
- System prompt — the base instruction: what the agent reads first.
- Guardrails — constraints: what must not be done.
- Memory — what to remember between steps.
- Tool guidance — how to work with tools: when to search the documentation and when to answer on its own.
- Subagents — delegation: one agent breaks down the task, others carry out parts of it.
- Skills — ready-made sets of techniques for typical tasks.
- Scripts — scripts for working with the environment.
- Plugins — middleware that changes behavior inside the tool-calling loop.
How the Two Levels of Recursion Work
The "²" in the name refers to two loops. Within a single task runs a Propose-Probe-Compose cycle. First, the agent solves the task with the base harness, and this reference run is recorded: what it read, which tools it called for no reason, what it forgot to check. Then the editor (by default Claude Opus 4.8, that is, a competitor's model) proposes two candidates: a general fix against typical weaknesses and a domain-specific one for a particular area such as legal procedures or spreadsheets. Each candidate is run on the same task, and the change in behavior is examined. The composer decides which edits to accept and assembles the final version.
An example from JobBench: the agent was assembling a set of documents with consistent figures. The first run scored 43.1: the numbers in the files did not match, and nobody reopened the final documents. The edits added source reconciliation and a mandatory re-check of the result, and the score rose to 77.8. This is one task, and it does not prove that the same edit will work everywhere.
The second loop operates between tasks. The reflector component analyzes the whole process and keeps only the lessons confirmed by comparing runs. They go into the improvement harness library, in the file EXPERIENCE.md. The next task starts from the base harness, and only the accumulated experience is carried over, which guides the editor on what to propose.
What the Benchmarks Showed
The authors tested the method on three task sets:
- WorkBuddy-Bench — office, web, and coding tasks. Peak: GLM-5 with OpenCode, from 64.2 to 75.7 points (+11.5).
- JobBench — long professional tasks close to a human's workday. Gemini 3.5 Flash with OpenCode: from 31.5 to 36.0 (+4.5).
- LAB from Harvey — a legal benchmark that measures the share of criteria passed. GLM-5: from 59.6% to 71.4% (+11.8 pp).
The most practical observation: the weaker executor gains more. After one optimization step, GLM-5 scores 75.7 (OpenCode) and 75.3 (Codex) on WorkBuddy-Bench, while Gemini 3.5 Flash with the standard harness scores 72.4 and 74.5. A weaker model with a tuned harness outperforms a stronger one with a standard harness.
Automation alone does not guarantee success. The competing ReasoningBank falls up to 6.6 points below the base harness on coding tasks, and Harness Scaling loses to the baseline in seven of twelve benchmark-configuration pairs. Harness² is above the baseline on WorkBuddy-Bench and JobBench in all four configurations.
Cost and Limitations
Each recursion step means additional runs, that is, tokens and time. Moving from a budget of k=1 to k=5 adds another 2.9 to 3.9 points on WorkBuddy-Bench, 3.0 to 4.2 on JobBench, and 2.6 to 6.7 pp on LAB, with diminishing returns. Moreover, a difference in behavior is not the same as an improvement: the system may lock in an edit that looks better but fails on the next task.
You can try it without paying for models: a local demo (Linux, Python 3.12, uv, the command harness2-demo --k 2) runs both levels with deterministic editors in about ten minutes. Real benchmarks require Vertex AI, gcloud, Node.js 20+, and either Docker or Pandoc with Poppler, and the runs are paid. This is a research framework, not a Google product: the API may change.
The practical takeaway for those building agents: before overpaying for a top model, measure how much systematic harness tuning gives you. All figures are taken from the authors' preprint and have not yet been independently verified.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What the Harness Consists Of?
The framework's README lists eight components, and Harness² can edit each of them: System prompt — the base instruction: what the agent reads first. Guardrails — constraints: what must not be done. Memory — what to remember between steps. Tool guidance — how to work with tools: when to search the documentation and when to answer on its own. Subagents — delegation: one agent breaks down the task, others carry out parts of it. Skills — ready-made sets of techniques for typical tasks.
How the Two Levels of Recursion Work?
The "²" in the name refers to two loops. Within a single task runs a Propose-Probe-Compose cycle. First, the agent solves the task with the base harness, and this reference run is recorded: what it read, which tools it called for no reason, what it forgot to check. Then the editor (by default Claude Opus 4.8, that is, a competitor's model) proposes two candidates: a general fix against typical weaknesses and a domain-specific one for a particular area such as legal procedures or spreadsheets.
What the Benchmarks Showed?
The authors tested the method on three task sets: WorkBuddy-Bench — office, web, and coding tasks. Peak: GLM-5 with OpenCode, from 64.2 to 75.7 points (+11.5). JobBench — long professional tasks close to a human's workday. Gemini 3.5 Flash with OpenCode: from 31.5 to 36.0 (+4.5). LAB from Harvey — a legal benchmark that measures the share of criteria passed. GLM-5: from 59.6% to 71.4% (+11.8 pp). The most practical observation: the weaker executor gains more.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.