--- title: "Google Harness2: AI Agents That Improve Without a Better Model" id: "342" type: "post" slug: "google-harness2" published_at: "2026-10-18T17:00:00+00:00" modified_at: "2026-10-18T17:00:00+00:00" url: "https://invetech.io/google-harness2/" markdown_url: "https://invetech.io/google-harness2.md" excerpt: "Google open-sourced Harness2, a framework that makes AI agents smarter by rewriting their instructions, not their code. Here is what that means for everyday users." taxonomy_category: - "Artificial Intelligence" - "Workplace Productivity" taxonomy_post_tag: - "AI" - "ai-agents" - "prompt-engineering" --- If you have ever wished your AI assistant would just get better at your specific job without you having to swap tools or pay for an upgrade, Google Harness2 is worth understanding. It is a new open-source framework that improves how an AI agent behaves, and it does that without touching the underlying model at all. That is a meaningful shift. Most people assume a smarter AI means a bigger, newer model. Harness2 challenges that assumption. ## What Google Harness2 Actually Does Think of an AI model as the engine inside a car. The engine itself never changes. But everything around it, the steering, the GPS, the dashboard settings, the route instructions, can be tuned. In AI terms, that surrounding layer is called a harness. It includes the system prompt (the standing instructions the AI follows), the tools it can use, its memory, and how it delegates subtasks. Google Harness2, released by Google Cloud AI Research and available on GitHub, rewrites those surrounding components automatically during use. According to the research team, model weights stay frozen. Only the harness changes. The framework exposes eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. Each one is mounted as a plain file, which means a developer can inspect and adjust them without deep machine-learning knowledge. Why does this matter to non-developers? Because the same principle applies to how you use any AI tool. The instructions you give, the context you provide, and the structure of your request are your personal harness. Improving those improves your results, no new model required. Our [Prompt Engineering and Context Engineering beginner’s guide](https://invetech.io/prompt-engineering-context-engineering-a-beginners-guide/) covers exactly that idea. ## How the Self-Improvement Loop Works Harness2 runs a two-level process the research team calls Propose, Probe, Compose. First, the agent attempts a task using its current harness. Then an editor proposes two alternative versions: one targeting general weaknesses like planning or tool use, and one targeting knowledge gaps specific to the current domain. Both alternatives run on the same task. A composer then compares the results and keeps only the changes that produced better observed behavior. After each task, a reflector stores the improved harness along with a lessons file. Future tasks draw on that library. Each new task still starts from the base harness, so a bad lesson on one task does not permanently break the agent. According to the Harness2 paper, one refinement step improved all four model and harness combinations tested across all three benchmarks. The biggest single gain was GLM-5 under OpenCode on WorkBuddy-Bench, which rose from 64.2 to 75.7, an 11.5-point improvement. The smallest gain was GLM-5 under Codex on JobBench, which moved from 23.6 to 24.8. The research team also found that weaker models benefited more than stronger ones. ## What This Looks Like in Real Work The self-improving harness concept is still primarily a developer tool. But the underlying logic maps directly onto problems small teams face every day. **A small law firm drafting client summaries** - **Problem:** The AI assistant produces summaries that are too formal and miss the plain-language details clients actually ask about. - **Usual approach:** The paralegal rewrites the output manually every time, which takes 20 to 30 minutes per document. - **With AI:** The paralegal refines the standing instructions (the prompt) after each document, noting what worked and what did not, building a personal lessons library. - **What to provide:** The task, the audience (clients with no legal background), the preferred format (short paragraphs, no Latin terms), and examples of past summaries that worked well. - **Benefit:** Each round of feedback makes the next draft closer to ready, reducing editing time. - **Limit:** The AI still needs human review for accuracy. Legal facts must be verified by a qualified person before anything goes to a client. **A solo retailer writing product descriptions** - **Problem:** Generic AI descriptions do not match the store’s voice and often miss the specific details shoppers care about. - **Usual approach:** The owner rewrites each description from scratch, which takes time away from other tasks. - **With AI:** The owner builds a short instruction file that describes the store’s tone, the typical customer, and the three details every description must include. That file becomes the harness for every new product. - **What to provide:** Brand voice notes, a sample description the owner likes, the product specs, and the target buyer. - **Benefit:** Consistent voice across all listings with far less rewriting. - **Limit:** The AI may still get product facts wrong. Every description needs a quick accuracy check before publishing. **A project manager summarizing weekly status reports** - **Problem:** Status reports from five team members arrive in different formats, and pulling them into one coherent update takes an hour every Friday. - **Usual approach:** Copy, paste, and manually reformat each section. - **With AI:** The manager creates a standing prompt that specifies the output format, the audience (senior leadership), and the three questions every summary must answer. The prompt improves each week based on what leadership actually asks. - **What to provide:** The raw reports, the output format, the audience, and any recurring questions from past meetings. - **Benefit:** The summary draft is ready in minutes, leaving time for the manager to add judgment and context. - **Limit:** The AI cannot know what happened in hallway conversations or what leadership is worried about this week. Human context is irreplaceable. ## The Cost of Waiting to Improve Your AI Setup Many people use AI tools the same way every week, with the same generic prompt they typed on day one. That is the equivalent of buying a new car and never adjusting the mirrors. The Harness2 research shows that even one round of deliberate refinement produced measurable gains across every test case. The framework is designed for what the research team calls the open world, meaning tasks that arrive without labels or pre-built test cases, which is exactly the situation most workers face. You do not need to run Harness2 yourself to apply this lesson. Every time you note what worked in an AI response and carry that forward into your next request, you are doing a manual version of the same process. The difference is that Harness2 automates and stores those lessons. Tools that help you think about how you are prompting AI, not just what you are asking, compound in value over time. Our article on [AI agents that speak first](https://invetech.io/ai-agents/) covers how leading teams are already building this kind of feedback loop into daily work. ## Limits and Responsible Use Harness2 is version 0.1.0, an early release. It runs on Linux with Python 3.12 or 3.13 and requires a Vertex AI project for model-backed runs, according to the GitHub repository. It is a developer tool, not a consumer product. Small business owners will not install it directly. The research team notes that the framework requires extra model runs per task, which adds cost and time. The gains on JobBench were modest in some cases, and one efficiency metric (steps per task) got worse for one setup on WorkBuddy-Bench. No framework improves every result in every situation. On privacy: Harness2 calls models through Vertex AI, Google’s cloud platform. As with any cloud AI tool, you should not send confidential client data, personal information, or proprietary business details to any external service without understanding that provider’s data handling policies. The same caution applies to ChatGPT, Copilot, and every other cloud-based AI tool. Human review of any AI output remains essential, especially in legal, financial, and medical contexts. For a broader look at how AI decision tools are evolving for small teams, the [Liquid AI d1-ultra decision model overview](https://invetech.io/liquid-ai-d1-ultra-decision-model/) offers a useful comparison point. ## The Takeaway: Better Instructions Beat Bigger Models Google Harness2 is a technical research release, but its core finding is practical and immediate. An AI agent improved by 11.5 benchmark points, according to the Harness2 paper, not because the model got smarter, but because the instructions around it got better. You already have that power. Every prompt you refine, every standing instruction you improve, and every lesson you carry forward from one task to the next is your version of recursive harnessing. The people and teams pulling the most value from AI tools right now are not the ones with the biggest models. They are the ones who treat their instructions as something worth improving. ## Your first step Open ChatGPT, Claude, or any AI assistant you already use. Pick one task you do every week, such as writing a summary or drafting a reply. Write a short standing instruction that names your audience, the format you want, and one thing the AI usually gets wrong. Save it as a text file and use it as your opening prompt next time. After two or three uses, update the instruction based on what still needs fixing. ## Sources - [Google Open-Sources Harness2, a Recursive Agent Harnessing Framework](https://www.marktechpost.com/2026/10/18/google-harness2-recursive-agent-harnessing-framework/) - [Harness2 GitHub Repository](https://github.com/google-research/harness2) - [Harness2 Research Paper (arXiv)](https://arxiv.org/abs/2610.14224) - [Harness2 Project Page](https://harness2.github.io/)