Google Research releases Harness², a framework that optimizes agent harnesses without retraining

What Harness² does

Google Research has published Harness², an open-source framework that treats everything around a language model — the system prompt, tool definitions, memory, subagents, skills and orchestration — as an optimizable system rather than something engineers tune by hand. The project's GitHub repository, which surfaced on Hacker News's front page, calls this surrounding machinery the "harness" and the act of improving it "harnessing."

The motivation stated in the repository is the open-world problem: agents regularly meet tasks that nobody anticipated when their harness was designed. Harness² responds by refining the harness automatically, and it does so while keeping model weights fixed — no fine-tuning is involved.

Two nested loops

According to the repository, the framework couples two levels of recursion.

The first operates at task level, through a cycle the authors name Propose–Probe–Compose. Candidate edits to the harness are proposed, then probed by actually running them on the current task, and a better harness is composed from the behavioral differences between those executions. Notably, this needs no ground-truth feedback — the loop learns from contrasts between runs rather than from labeled answers.

The second level persists across tasks. After each one, the harness that was adopted, along with lessons backed by execution evidence, feeds into a lasting improvement harness that guides future refinement, so the system accumulates experience instead of starting over.

Eight components are exposed for editing: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Recursion can run in parallel mode, which probes many candidates at once and merges their evidence, or sequential mode, which carries a refined harness through successive steps. Both OpenCode and Codex executors are supported.

Reported results

The maintainers evaluated the framework on three professional benchmarks — JobBench for professional work, LAB for legal work, and WorkBuddy-Bench covering office, web and code tasks — across four model–harness configurations. They report that a single Harness² step lifts GLM-5 and Gemini 3.5 Flash above their base harnesses, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate. Increasing the task-level recursion budget, controlled by the --k flag, is said to bring further gains.

Practical notes

The project targets Linux with Python 3.12 and uv; model-backed runs also require Node.js 20+, the Google Cloud CLI and Vertex AI configuration. A local demo runs both recursion modes with deterministic editors and no model credentials, while the benchmarks add dependencies such as Docker for WorkBuddy-Bench, and Pandoc and Poppler for LAB. API costs apply to model-backed runs, and the documentation advises starting with a small task selection.

One caveat is flagged prominently: executor network access is unrestricted and Codex runs with its sandbox disabled, while public benchmark rubrics may be reachable online. Users are told to configure network restrictions before evaluation so agents cannot fetch answers or grading criteria.

The code is Apache 2.0 licensed, with the standard research disclaimer that it is not an officially supported Google product. An accompanying paper is cited as arXiv preprint 2610.14224, authored by Ruiyao Xu and colleagues.

Why it matters

Agent engineering has quietly become a scaffolding craft: teams spend weeks hand-tuning prompts, tool descriptions and memory policies, and that work silently expires whenever the task distribution shifts. Harness² promotes this layer to a first-class optimization target that improves itself from its own executions, with no retraining and no labeled feedback. If the reported gains hold up outside the authors' own benchmarks, the practical implication is significant — better agent performance could come from the harness loop rather than the next model release, and existing models could be stretched further on unfamiliar work. As with any self-reported results from a research repo, independent replication is the next test.