Editorial illustration for Google Open-Sources Harness2, a Recursive Agent Framework With 8 File Components
Google Open-Sources Harness2 Agent Framework
Google Cloud AI Research released Harness² on GitHub this week, versioned at 0.1.0, and it skips the part everyone expects from an agent framework: no trained parameters, no fine-tuning run, no new model weights at all. Instead it works on the 8 files sitting around a frozen model, things like system prompts, memory, tool guidance and plugins, and rewrites them during deployment to make the agent perform better on tasks it has never seen before.
That's the core bet behind the project. Most self-improvement research leans on labeled data or test cases to tell the system whether it's getting closer to a right answer. Harness² assumes none of that exists, because in real deployment it usually doesn't.
Tasks show up unannounced, there's no verifier grading the output, and the model itself stays exactly as it was shipped. What changes is the scaffolding: the prompts, the subagents, the scripts that route a request to the right tool.
Google's team tested the approach across benchmarks, running models like GLM-5 through harnesses like OpenCode and Codex to see what a single refinement pass could do without touching the underlying weights.
Google Cloud AI Research has open-sourced Harness². It is a recursive agent harnessing framework that improves the code and files around a frozen model.
Why this matters
Harness² is a bet that the next round of agent gains comes from scaffolding, not scale. Google Cloud AI Research is handing developers 8 editable files (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins) and saying: tune these around a frozen model and watch performance move. That's a cheaper lever than waiting on a model upgrade, and it fits where a lot of teams already are, stuck with budget caps on Vertex AI calls and no appetite for fine-tuning.
We'd flag two things before anyone ports production agents over. Version 0.1.0 and zero disclosed benchmarks beyond "1 refinement" is thin evidence for a framework asking you to restructure your tool loop. And it's Vertex AI-bound on Linux with Python 3.12/3.13 through uv, so this isn't a drop-in for teams running elsewhere.
Worth watching: whether Google publishes real before-and-after numbers on specific tasks, and whether the 8-file layout gets adopted outside Google's own tooling or stays a curiosity next to the 7-file setups it's extending.
Related Reading
Browse all Google news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives Meta turns Mac into AI supercomputer; gender gaps show in training, not access
From Archives DeepSeek releases V4.1-Flash-1017, scores 77.0 on DeepSWE, opens Search API
From Archives Google Gemini 4 Argon scores 91.3% on ARC-AGI-2, nearing AI frontier
Research & Benchmarks
From Archives Nvidia Says Nemotron 3.5 Super VL Outperforms Gemini Ultra in Agent Tests
Common Questions Answered
What are the 8 file components that Harness2 uses to improve agent performance?
Harness2 works with eight editable files surrounding a frozen model: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. These files are rewritten during deployment to optimize agent performance on new tasks without requiring any model retraining or parameter updates.
How does Harness2 differ from traditional agent frameworks that require fine-tuning?
Unlike conventional agent frameworks that rely on trained parameters, fine-tuning runs, and new model weights, Harness2 operates entirely on the scaffolding around a frozen model. This approach eliminates the need for model training while still achieving performance improvements through strategic modifications to the surrounding files and configurations.
Why is Harness2's focus on scaffolding instead of scale significant for development teams?
Harness2 represents a bet that agent performance gains come from optimizing scaffolding rather than scaling models, which is a cheaper and more accessible lever for teams with budget constraints on API calls. This approach fits teams already stuck with budget caps on Vertex AI calls who lack the resources for expensive model upgrades or fine-tuning operations.
What does it mean that Harness2 is a recursive agent harnessing framework?
Harness2's recursive nature refers to its ability to iteratively rewrite and improve the 8 file components around a frozen model to enhance agent performance on tasks it has never encountered before. This recursive improvement process allows the framework to adapt and optimize without modifying the underlying model itself.
