Scientists Find a Better Way to Train AI That Manages Other AI

Imagine a project manager coordinating engineers, designers, and writers. The manager needs to choose who does what, but also give detailed instructions. Training an AI to do this job used to be messy: the AI learned both the 'what' and the 'how' at the same time, and those lessons often clashed. A new method called Harness-RL fixes that by separating the two — the AI learns simple decisions (like which task to assign) and detailed execution (like exactly how to phrase a request) as two separate skills.

The method also solves a practical puzzle: when the 'boss' AI directs other AI specialists, it often can't see inside them. Think of it like managing outside contractors — you can give them orders and see the final result, but you can't watch their every move. Harness-RL lets the boss learn from those black-box interactions by tracking records of calls and building a tree of how the work branched out, even when tasks run in parallel or contexts get rewritten along the way.

In tests, this approach made AI 'bosses' significantly better at coordinating teams on seven challenging benchmarks involving multi-step questions and research-style retrieval. Even more impressive, it worked well with relatively small AI models (around 1.5 to 3 billion parameters), which are far cheaper to run than the giants behind many AI products. That means the technique could make smart, cost-effective AI assistants available to more people and businesses.

The catch: this is research, not a product yet. It also works best when only the boss is trained and the specialist sub-agents stay fixed — which is actually a common real-world setup. Still, by tackling one of the trickiest parts of AI teamwork, Harness-RL brings us closer to AI systems that can reliably handle big, multi-step jobs without hitting a wall.