The wrapper around your AI agent may be as trainable as the model inside it.
Researchers have published a paper proposing that the execution harness surrounding a large language model — the scaffolding that sequences calls, handles retries, and structures verification steps — should be treated as a learnable control layer rather than fixed infrastructure. They formalize this as a "Harness MDP," a decision-making framework where a lightweight controller picks structural execution actions while the underlying LLM stays frozen. That controller is trained from recorded rollouts using a technique called advantage-weighted regression, rewarded only at task completion. Across six controlled domains and two public benchmarks, the approach consistently improved verification behavior and, in several cases, final task quality — with the strongest gains on retail task automation, database benchmarking, and coding tasks paired with a structural verifier.
Most agent improvement work targets the obvious levers: swap the model, tweak the prompt, rewrite the workflow. This paper pushes in a different direction by leaving the LLM untouched and optimizing the frame around it — which matters because deployed models are often frozen for cost or compliance reasons. The authors also introduce a "Harness Maturity Score" that separates process quality from outcome quality, a useful distinction that standard benchmarks tend to collapse into a single accuracy number.
The caveat is real and the paper is upfront about it: better process control only translates to better final answers when the offline training data already contains high-quality examples to learn from. That limits how far this approach can generalize without richer replay buffers — and it means the method is more "learn to follow good patterns" than "discover new ones."