A new research framework treats agent improvement as a balancing act between goals, and the security gains hold up on tasks it never saw
Researchers from Google Cloud AI Research, Google DeepMind, and Carnegie Mellon University have introduced Harmony Harness, a framework for automatically evolving the scaffolding that surrounds large language model agents.
The paper was submitted to arXiv on October 10, 2026, under the identifier 2610.15150.
If a large language model is the engine, the harness is everything else that turns it into a working car: the surrounding structure that lets the model actually get things done. Previous harness engineering was largely manual, with 2026 seeing quick adaptation to automated optimization methods.
Harmony Harness pushes that automation in a specific direction. Instead of optimizing one blended score, it frames harness evolution as a multi-objective optimization problem. Security and functionality get treated as distinct targets rather than mashed into a single number.
To manage those competing goals, the framework uses decomposition-based evolutionary search. It maintains what is called a Pareto frontier of candidate solutions, where each solution is the best available option for some particular tradeoff—you cannot improve it on one goal without giving something up on another.
The second key ingredient is what the paper calls factorized evolution operators. These let the framework refine general-purpose mechanisms separately from task-specific adaptations. The pieces can then be combined dynamically at inference time, when the agent is actually running.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
On two benchmarks, SecRepoBench and BaxBench, Harmony Harness scored more than 10 percentage points higher than prior strong single-objective baselines. Both benchmarks focus on whether AI agents write code that is secure, not merely code that runs.
Harnesses evolved under the new framework retained their security benefits more effectively on held-out and unseen tasks. Traditional approaches that do not explicitly handle multiple objectives tend to show overfitting and regressions, according to the research.
The research also reports that explicitly handling multiple objectives improves both security and functional performance on out-of-distribution tasks.
The “agent harness” concept gained traction in industry conversations across 2025 and 2026. Related efforts such as Harness2 and EvoHarness-RL emerged around the same time, pointing to growing research interest in agent scaffolding and security.
For teams building AI agents, if security is folded into a single performance score, the optimizer may trade it away when it is convenient. Keeping security as its own objective appears to make those gains stickier, at least on the benchmarks tested here.
As of October 15, 2026, no commercial product details had been announced, and the work carries no announced integrations. The involvement of both Google Cloud AI Research and Google DeepMind suggests the problem is being taken seriously at multiple levels of the company.
Benchmark gains do not always translate cleanly into deployed systems, and the results here center on secure-code tasks specifically. With Harness2 and EvoHarness-RL appearing alongside Harmony Harness, automated harness evolution looks set to become a crowded research area.