As multi-agent workloads grow in depth and complexity, one question becomes central: How do these systems decide what to do? Our latest Kite AInsights article by our Researcher Kevin Ros, explores the frameworks behind that process. As tasks grow more complex and agents become more specialized, planning determines whether systems can translate broad goals into coordinated action. The question is no longer “Can agents solve tasks?” It’s “Can they decompose, delegate, and execute effectively?” At the core of this work is Agent-Oriented Planning (AOP), guided by three decomposition principles: ☑️ Solvability: each subtask can be handled by a single agent ☑️ Completeness: tasks fully cover the user’s goal ☑️ Non-Redundancy: no duplicates or irrelevant work The meta-agent plays the coordinating role interpreting the query, distributing tasks, and refining the plan through: • A Detector to catch missing or redundant subtasks • A Reward Model to evaluate subtask solvability The result? AOP delivers ~10% higher accuracy vs. single-agent baselines, and ~4% over naive multi-agent setups. Planning is becoming the connective tissue for multi-agent systems - defining how they structure, coordinate, and complete complex workflows.