Kite AI @GoKiteAI shared a pretty interesting study: How multi-agent systems (MAS) learn to 'plan.' Simply put, it's about how AI agents can divide tasks and collaborate more intelligently. They introduced a new method: Agent-Oriented Planning (AOP). It enables the meta-agent (main control agent) to break down tasks, allocate them, and integrate results more effectively. Because right now, many multi-agent systems have a common issue: Tasks are broken down too much, assigned chaotically, and often duplicated. AOP is designed to solve this. AOP has three core principles: Solvability: Each sub-task can be completed independently; Completeness: All sub-tasks combined can solve the original problem; Non-Redundancy: No duplication, no excess. Under these three rules, the AOP process looks like this: → User asks a question → Meta-Agent breaks down the task → Detector checks for duplicates or omissions → Reward Model predicts feasibility → Individual agents execute tasks → Continuous feedback and optimization. In their experiments, the research team used GPT-4o as the meta-agent to coordinate four types of sub-agents: math, search, code, and common sense. Results showed: Accuracy was 10% higher than single-agent systems, 4% higher than regular multi-agent systems. Costs were slightly higher, but the outcomes were more stable. Future directions are also clear: 1️⃣ Enable agent-to-agent collaboration; 2️⃣ Introduce human-in-the-loop participation; 3️⃣ Build trust and verification mechanisms, Especially for scenarios where agents spend real money, like buying plane tickets or paying for services. Finally, Kite AI also shared their vision: To build an infrastructure where agents can truly act independently and be trusted. This includes identity systems, payments, governance, and verification. Essentially, they're laying the groundwork for the entire AI Agent economy.