New Google Tool Lets AI Agents Self-Optimize Prompts

The open-source framework allows models to iterate on their own performance without the need for full retraining.

Updated on Oct. 21, 2026 in Artificial Intelligence

Isometric editorial illustration showing a stack of multi-faceted geometric glass prisms, representing modular AI system optimization. — Google researchers have released Harness2, an open-source tool enabling AI agents to self-optimize configurations and improve performance without requiring full model retraining. AI Illustration. Upload story photo >

Live Poll

Do you trust AI agents that can rewrite their own instructions without human intervention?

Google researchers have open-sourced Harness2, an agentic tool that enables AI to modify its own system prompts, tools, and memory without requiring model retraining. This research-stage framework utilizes a loop to iterate on performance, though Google notes it is not an officially supported product.

Why it matters

Harness2 addresses the high computational costs of fine-tuning by allowing agents to self-optimize their operational parameters. It represents an shift toward modular, iterative improvement for AI systems in resource-constrained environments.

The tool uses a Propose, Probe, and Compose loop to test harness changes against an editor model, specifically Anthropic's Claude Opus 4.8. It achieved a 75.6 score on the legal LAB test compared to the 76.5 score recorded by the Meta-Harness benchmark.

The players

Google

A global technology company focused on search, cloud infrastructure, and large-scale machine learning research.

Anthropic

An AI research and product company that develops the Claude series of large language models.

The details

Harness2 manages an AI agent's behavior through eight editable files, including guardrails and subagents, which are stored in the system. The agent executes a task, suggests two different versions of its configuration, and uses an editor model to evaluate which version yields superior performance. Lessons learned from these iterations are recorded in an EXPERIENCE.md file, allowing the system to refine its prompts and tools over time without full retraining. The tool requires a Linux environment with Python 3.12 or 3.13 and active credentials for a Google Cloud account.

October 9, 2026: The Harness2 research paper was submitted.
October 21, 2026: The article regarding Harness2 was published.

The Tech Race

Harness2 enters a competitive landscape defined by efforts to improve agentic reasoning without massive re-training cycles. It operates as a research alternative to benchmarks like Meta-Harness, aiming to solve efficiency bottlenecks in current AI agent architectures.

Developers can implement Harness2 today by accessing the repository under the Apache-2.0 license. Users must have a Linux-based environment and a Google Cloud account to leverage the framework's self-optimizing loops.

The takeaway

Harness2 provides a modular path for agents to improve their own logic without the overhead of expensive fine-tuning. Future updates to monitor include the release of more performance benchmarks on the tool's GitHub repository.