Google Research has released Harness², an open-source tool that improves AI agent performance without retraining the underlying model. When an agent fails a task, the system diagnoses what went wrong and adjusts how the agent operates rather than touching the model itself.
Why this matters for teams building on AI agents
The standard path to a smarter agent is fine-tuning the model — expensive, slow and dependent on labelled data. Harness² leaves the model untouched and works on the layer around it instead.
Four things it can change:
- instructions — prompts and system messages the agent receives
- settings — runtime parameters and configuration
- memory — what the agent retains between steps
- tools — the function set and how those functions are called
If an agent handles a request poorly, Harness² tests alternative instructions or finds a different way to call the available tools. It is closer to automated prompt engineering than to machine learning in the classical sense.
How the loop runs
The system works iteratively: it runs the agent, scores the result, generates hypotheses about what to change and tests them. The cycle continues until quality improves or the attempt budget runs out.
For teams running automation on top of AI agents, that means less manual tuning. Instead of rewriting prompts by hand for each new task, the system searches for a working configuration on its own.
Open source
Harness² is published as open source, so it can be integrated into existing pipelines, adapted to specific tasks and run without a dependency on Google's cloud API. Teams working on agents for data scraping, media buying, lead processing or reporting automation can test the tool against their own data at no additional cost.
