Skills
Give the orchestrator a price list: knowing each executor's capability and cost changes the plan, not just the routing
EASy trains an orchestrator that holds explicit knowledge of its executors' capabilities and costs, then runs a milestone-plan-act workflow that decomposes tasks into steps on a dependency-aware execution graph. Training uses a tree-structured rollout to explore alternative decompositions and plans, rewarded on task correctness, execution efficiency, and trajectory completeness jointly. Across mathematical reasoning, embodied decision-making, and deep research benchmarks it reported better performance-efficiency tradeoffs than strong agentic baselines — useful if your multi-agent system optimizes success rate and ignores what each step costs.
↳ Follow the thread