A new framework lets AI agent teams decide, mid-task, how many agents they need and who answers to whom.
Researchers describe DHCG, short for Dynamic Hierarchical Collaboration Graph, a framework that builds its own org chart for AI agent teams as they work. Three modules run the show: a Planner reads the query and the system's own progress, a Worker executes roles the Planner assigns, and a Generator produces outputs along the way. The Planner can spin up new specialist roles mid-task, route relevant information to each one, and decide whether to wrap up early or expand further. The team trained the Planner with a technique called action-aware preference optimization, aimed at making better calls about when to grow or shrink the group.
Tested on code generation, math, and domain-specific reasoning benchmarks, DHCG beat a single-agent baseline by just over 13 points, and beat other multi-agent setups, both fixed and dynamic, by roughly 3 to 8 points. That matters because most multi-agent LLM systems today use a fixed cast of agents or add agents crudely, without much sense of when a new role actually helps. DHCG's premise is that the right team shape depends on the problem and should shift as the system learns more about it, closer to a manager adding a specialist only once a blocker shows up.
The comparisons are against other multi-agent baselines, not against a single well-prompted model given more tokens and retries, so it is still an open question how much of the gain is genuine coordination versus just more compute dressed up as teamwork.