# Harness2 Harness² is a research framework from [[Google]] Research that improves the harness around a model while the model's weights stay frozen. The README defines a harness as "the system around a model that enables it to act as an agent". Harnessing is the work of optimizing that system. Code and paper came out in October 2026 (arXiv 2610.14224, 9 October), under Apache 2.0. Despite the org name, the repo says "This is not an officially supported Google product." The problem it goes after: an [[AI Agent Harness]] is designed for the tasks its authors had in mind, and an agent in the open world keeps running into tasks nobody planned for. Harness² adapts the harness to each new task, then keeps whatever worked for the next one. ## Two levels of recursion - **Task-level.** A contrastive Propose-Probe-Compose step. An editor proposes edits to the harness (generic ones and domain-specific ones), probes run each candidate on the same task, and a composer builds a refined harness from how the runs behaved differently. No ground-truth verifier is needed. You scale it in parallel (`--k K --mode parallel`: K generic and K domain candidates, K compositions) or sequentially (K refinement steps in a row) - **Harnessing-level.** A persistent "improvement harness". After each task it keeps the adopted harness plus the lessons the evidence supports, and uses them to steer the next refinement The editable surface has eight parts: system prompt, [[AI Guardrails|guardrails]], [[AI Agent Memory|memory]], tool guidance, [[AI Subagents|subagents]], [[AI Agent Skills|skills]], scripts and plugins. So it rewrites code-like artifacts too, not just prompts. The agent doing the work is [[OpenCode]] by default, or [[Codex CLI|Codex]] (`--substrate cx`). Codex runs with its sandbox disabled and the executor has unrestricted network access, so keep it away from anything you care about. ## Numbers All from the authors. One Harness² step improves GLM-5 and [[Gemini 3.5 Flash]] over their base harnesses on three professional benchmarks: up to +11.5 points on WorkBuddy-Bench (office, web and code), +4.5 on JobBench and +11.8 percentage points on LAB's partial pass rate (legal work). A bigger task-level budget gives more. The README only reports gains, never absolute scores. There's a credential-free demo (`harness2-demo --output runs/demo --k 2`) with deterministic editors. Real runs need Linux, Python 3.12, uv, Node 20+ and [[Vertex AI]], and they cost API money. The README also warns that benchmark rubrics are public, so you have to block the agent from looking up the answers during evaluation. ## Reception on Hacker News The repo got 81 points and 19 comments (around 2026-10-20). The main pushback was "we already do this". mchusma and jmathai said that learning from past runs to update skills and system prompts is routine, and plenty of people automate it alongside code changes. cykros said it sounds like what [[Hermes Agent]] already does. endangeredhuman pointed at Google's move from the open-source [[Gemini CLI]] to the closed [[Antigravity CLI]]: "If the harness is closed, why open the 'harnessing'?" ridgeComet617 missed "a single end to end example you can actually run", and several commenters mocked the AI-generated README image. ## My take The HN crowd is half right. Updating skills and prompts after a run is what any serious [[Harness Engineering]] setup already does, mine included. What's new here is the probe step: running competing harness edits on the same task and comparing behavior, without a verifier. That's a partial answer to the question nobody answered in the July harness engineering thread: how do you know a harness change actually helped? I wouldn't run it as-is. It's research code tied to Vertex AI, with the sandbox off and only relative gains reported. The pattern is worth stealing, though. Before adopting a skill or prompt change, probe it against the current version on the same task. ## References - Source code: https://github.com/google-research/harness2 - Paper: [Harness²: Recursive Agent Harnessing for an Open World (arXiv 2610.14224)](https://arxiv.org/abs/2610.14224) - Website: https://harness2.github.io - Hacker News discussion: https://news.ycombinator.com/item?id=50151059 ## Related - [[Harness Engineering]] - [[AI Agent Harness]] - [[Hermes Agent]] - [[AI Agent Skills]] - [[AI Agent Memory]] - [[Context Engineering]] - [[OpenCode]] - [[Codex CLI]] - [[Gemini CLI]] - [[Antigravity CLI]] - [[Google]] - [[Hermes Agent Self-Evolution]] - [[GEPA]]