Multi-agent AI systems for chip design are evolving from rigid, top-down hierarchies into flatter, more dynamic organizations. Experts from major EDA vendors, speaking at a Design Automation Conference roundtable, compared the new structure to an orchestra: one conductor, but groups that can play individually or together, with guardrails and a shared score. Some systems are even "headless," with no central master agent.
The shift reflects a practical need for fluidity. Agents are assembled dynamically based on the task, and an orchestrator layer calls on resources as needed. Importantly, agentic capabilities will increasingly come from deterministic physical engines, not just AI models. These agents will talk to other agents in different formats to help ensure accuracy that AI models alone cannot achieve.
The roundtable also stressed that context management matters more than agent count. Deploying many agents is useless if they all share the same local context. The power comes from managing which hypotheses each agent generates and where that context comes from, and from delivering the full flow rather than point acceleration around one tool. Human effort remains necessary for copyrightability, and on-prem proprietary models can protect prompt IQ and IP.