Log in Sign up
Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time News Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size AI Agents Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains Apple Intelligence Apple Rewrites App Store Rules to Purge AI Nudify Apps, Putting Developer Accounts on the Line Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports Opinion Taste-Bench From Microsoft and CityU Shows AI Agents Pick the Right Long-Term Move Only 60 Percent of the Time News Kyutai Releases REKI, a 2.4B-Parameter Robot Model With Built-In Memory That Punches Above Its Size AI Agents Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains Apple Intelligence Apple Rewrites App Store Rules to Purge AI Nudify Apps, Putting Developer Accounts on the Line Opinion He Sold Everything and Rode a Bus 2,000 km to Marry an AI Farm Girl With 300,000 Followers Business Anthropic Pledges 5 Percent of Itself to a Public Interest Fund That Grows With Its Valuation AI Agents NVIDIA Open-Sources ASPIRE, the Robot Framework That Turns Failed Grasps Into a Growing Skill Library Coding Assistants PostgreSQL 19 Slips to November 19 After LLM-Assisted Code Review Turned Up Forty Defect Reports

Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains

AI Agents · By: AIReport Admin · 1 hour ago · 👁 6
Google Open-Sources Harness2, a Framework That Lets an Agent Rewrite Its Own Scaffolding for Double-Digit Gains

Google Cloud AI Research released Harness2, an Apache 2.0 framework that optimizes eight editable components around a frozen model through a Propose-Probe-Compose loop, with no labels or external verifier required. Google-reported numbers show one refinement step lifting GLM-5 from 64.2 to 75.7 on WorkBuddy-Bench.

Google Cloud AI Research has open-sourced Harness2, versioned 0.1.0 under an Apache 2.0 license on GitHub and described in a paper on arXiv (2610.14224), and its pitch inverts the usual scaling story: your agent may get better without a better model. The framework trains zero parameters. Instead it recursively rewrites the harness, meaning everything that surrounds a frozen model during deployment, from system prompt to plugins.

The harness in Harness2 is exposed as eight editable components, each mounted as a file: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins acting as middleware. Together they govern context construction, the tool interface, delegation and behavior inside the tool loop, extending the seven-file layout of earlier agentic harness engineering work by one component.

The mechanism is a two-level recursion built for the hardest setting: open-world tasks that arrive unseen, where no verifier confirms whether an answer is correct. At the task level, a contrastive Propose-Probe-Compose step runs the task once with the base harness, has an editor propose two candidates, one invariant targeting general agent failures like planning and tool use, one domain-specific, then probes both on the same task and keeps only edits backed by observed behavioral contrasts. No ground truth is ever consulted, because in real deployment there usually is none.

The second level accumulates. After each task a reflector stores the adopted harness plus an EXPERIENCE.md file of lessons, and that library steers future proposals, with each new task still starting from the base harness. Google reported that removing the library cut WorkBuddy-Bench performance from 77.5 to 74.4 in its own ablation, the clearest internal evidence the accumulated lessons do work.

The headline numbers are Google-reported and have not been independently verified. One refinement step improved all four tested model-and-harness setups on all three benchmarks, with GLM-5 under OpenCode rising from 64.2 to 75.7 on WorkBuddy-Bench, a gain of 11.5 points, and Gemini 3.5 Flash reaching 80.8 at five steps, up from 72.4 at base. Gains on JobBench and on the LAB partial-pass metric were smaller. Executors tested were OpenCode and Codex, with Claude Opus 4.8 as the default editor model, and results are means over three runs.

The context here is a discipline that has been hardening all year under the name harness engineering, the observation that the same model can swing dramatically in measured capability depending on the scaffolding wrapped around it. Harness2 turns that intuition into an automatically tunable artifact and treats agent capability as an improvement axis separate from model capability, one that is usually far cheaper than fine-tuning because it never touches the weights.

The caveats are the ones Google discloses and a few it does not. Contrastive selection improves measured behavior but does not guarantee correctness, so independent verification remains necessary for legal and safety-critical tasks. The project currently targets Linux with Python 3.12 or 3.13 and is documented against Google's own tooling. And this is a 0.1.0 release whose evidence base is one paper's benchmarks. Even so, the direction is clear: as agent platforms commoditize, the next round of gains may come less from training runs and more from giving the agent the tools to edit its own working environment, safely and on a budget.

Comments (0)

Log in to join the discussion

Log In

No comments yet