Your AI agent is only as good as its operating system, yet we are still building these systems by hand. Researchers have introduced JIT-Agent, a model that automatically designs and evolves the complex rules governing how AI agents think, plan, and act. This breakthrough suggests that the secret to superhuman AI performance might not lie in bigger models, but in smarter, self-improving frameworks.

Think of an agent harness as the brain’s workflow software. It manages memory, decides on planning strategies, and orchestrates tools. Currently, engineers manually craft these workflows for specific tasks, a process that is rigid and impossible to scale. JIT-Agent changes this by treating the harness as a composable, machine-generatable artifact. It acts as a helper that synthesizes task-adaptive rules on the fly, repairs broken workflows during execution, and learns from a growing archive of past successes to evolve its own strategies.

The results are startling. When paired with existing models, JIT-Agent dramatically boosts performance. DeepSeek-V4-Flash, equipped with this new harness intelligence, outperformed GPT-5.6 on major benchmarks like DeepSearchQA (+9.1) and OdysseyBench (+4.3). Even stronger models like GLM-5.2 saw gains of up to +20.2 points. Crucially, these JIT-generated harnesses compete with mature, hand-tuned agent runtimes like OpenCode and Claude Code, proving that the framework itself is a powerful, trainable dimension of capability.

This research establishes harness intelligence as a compounding force orthogonal to model scaling. We no longer need to wait for larger foundation models to solve harder problems. By focusing on just-in-time harness generation, we can unlock superior performance from off-the-shelf AI, turning static agents into dynamic, self-optimizing systems. The future of AI lies not just in what it knows, but in how it organizes its knowledge in real time.