Multi-agent systems can improve performance by dynamically adapting their internal communication structure at inference time, rather than relying on static pre-designed topologies.
MANTA is a framework that lets multi-agent AI systems automatically reorganize how they communicate and work together during execution. Instead of fixing agent roles and communication patterns upfront, MANTA monitors how agents collaborate and adjusts the team structure in real-time when needed—changing who talks to whom, agent responsibilities, and validation steps.