A new technique lets AI systems decide what needs doing before locking in who does it.
Researchers behind a method called DeOrch split multi-agent AI orchestration into two separate stages. First, a planner breaks a task into subtasks without knowing anything about which AI models or tools, called workers, are available to do them. Only afterward does a separate matcher pick workers, using a lightweight scoring system that is tested against a fixed probe set and refined with a contextual bandit algorithm as it runs. In tests across a range of tasks, including ones unlike anything in its training data, DeOrch beat earlier automated orchestration methods while making fewer calls to worker agents, and it kept working when swapped onto an entirely new set of workers without retraining.
Most multi-agent AI frameworks today bake planning and worker choice together, which means swapping in a new model or tool can require retraining the whole system. DeOrch's split approach makes the planner reusable across worker pools, which matters for anyone building agent systems that need to add new tools or upgrade models without starting over. It also untangles credit assignment: it is clearer whether a bad outcome stems from a bad plan or a bad worker pick.
That is the kind of modular plumbing AI agent tooling has needed for a while - and it is still a research paper, not a shipped framework.