# Google releases Harness² as open source, letting an agent rewrite its own prompts, tools and memory at run time without retraining the model

A separate paper, HybridAgent, reports 35.5% to 54.7% fewer tokens by having the model reason in latent space and switch to text only when it needs to call a tool.

- Published: 2026-10-19T06:14:35.233Z
- Canonical: https://polylog.news/ai/2026-10-19/google-releases-harness-as-open-source-letting-an-agent-rewr
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Hacker News (arXiv)](https://arxiv.org/abs/2610.18861), [Hugging Face Papers](https://huggingface.co/papers/2610.14224), [Pasquale Pillitteri](https://pasqualepillitteri.it/en/news/25395/google-harness2-ai-agents-improve-themselves), [AI Weekly](https://aiweekly.co/alerts/hybridagent-toggles-latent-reasoning-with-tools-cuts-cost-37x)

Google Cloud AI Research has released Harness², an open-source framework that improves an agent's harness while the agent is working. The harness is everything around the weights: system prompt, tools, memory, skills and subagents. [The Russian-language channel AI ML Big Data](https://t.me/ai_machinelearning_big_data/11155) summarized the premise: an agent's results depend on that environment as well as on the model, and today a person designs it once and freezes it.

The method runs two loops, according to the [paper page](https://huggingface.co/papers/2610.14224) and a [technical summary](https://pasqualepillitteri.it/en/news/25395/google-harness2-ai-agents-improve-themselves). At the task level, the system proposes several harness edits, tests them against each other and keeps the best combination. At the harness level, lessons from those edits update a second, persistent harness that the agent carries into later tasks. The method does not need a ground-truth verifier. Most self-improvement loops depend on verifiable rewards such as unit tests or exact-match answers, so this widens the set of tasks where the approach applies. The authors report gains for GLM-5 and Gemini 3.5 Flash in every configuration they tested on WorkBuddy-Bench, JobBench and LAB. Running five candidate edits in parallel added 2.9 to 3.9 points on WorkBuddy-Bench.

A second paper targets agent cost at the reasoning level. [Hybrid Reasoning Agents](https://arxiv.org/abs/2610.18861), which drew discussion on Hacker News, lets a model think in continuous latent space and switch to text only when it must call a tool. A lightweight module called HybridLink handles transitions between the two modes, and a continuous working memory keeps hybrid context across turns. Training runs in two stages: HybridSFT stabilizes tool-integrated generation, then HybridRL teaches the policy when to think silently and when to act. Across 11 benchmarks at three model scales, the authors report a 10.4% performance improvement over text and latent baselines, 35.5% to 54.7% fewer tokens, 1.8x to 2.6x faster inference and 2.7x to 3.7x lower cost, [per AI Weekly's summary](https://aiweekly.co/alerts/hybridagent-toggles-latent-reasoning-with-tools-cuts-cost-37x).

All of these figures come from the authors, and no one has reproduced them independently yet. Harness² results come from benchmarks the authors chose, and a harness tuned per task can overfit to what an evaluator rewards. HybridAgent's savings come with a cost to oversight: reasoning done in latent space produces no readable chain of thought for a monitor to inspect.

## What this means

Both papers get more performance from fixed weights. Harness² does it through configuration search, and HybridAgent does it by spending fewer decoded tokens. Both shift value from pretraining scale toward the scaffold and the post-training recipe. The beneficiaries are mid-tier and open-weight models such as GLM-5 and Gemini Flash-class models, along with teams that cannot afford frontier API calls for every agent step. The labs most exposed are those whose pricing assumes customers need the largest model for agentic reliability. The consensus assumes agent quality mainly follows model size. These results suggest configuration and reasoning format account for a measurable share of it. The desk's read is that within months, harness-optimization loops of this kind will appear in mainstream agent frameworks as an automated stage, because the code is open and needs no retraining. That depends on independent groups reproducing gains on benchmarks the authors did not choose. The read would be wrong if those reproductions show the gains disappear outside WorkBuddy-Bench-style tasks. For latent reasoning, the deciding question is whether safety teams accept the loss of a readable trace. If they do not, labs will restrict it to low-risk tool loops.

## What to watch

- Whether major agent frameworks build Harness²-style automatic harness editing into their tools. That would show the method moving from paper to production.
- Independent reproductions of HybridAgent's token savings on public agent benchmarks. Author-reported efficiency gains often shrink when other groups test them.
- Statements from labs or regulators on whether latent-space reasoning is acceptable in monitored deployments. They decide whether the cost savings can be used where oversight is required.