MANTA: Dynamic AI Communication Networks

Static communication graphs are holding back multi-agent systems. According to new research published on arXiv by Mao-xun Huang, Jerry Wang, Yi-Cheng Lai, Zhengxin Zhang, Claire Cardie, and collaborators, fixed topologies create rigid bottlenecks during real-time problem solving.

Their framework, MANTA, introduces Multi-Agent Network Topology Adaptation to let collaboration structures self-evolve during deployment.

Inference-Time Structural Evolution

Existing multi-agent frameworks treat communication graphs as static choices or offline tuning targets. MANTA shifts this paradigm by initializing a task-conditioned topology from prior structural experience before execution begins.

During active execution, the system continuously tracks collaboration traces.

When an active organizational layout proves insufficient, MANTA executes targeted, bounded structural modifications. These updates alter agent roles, communication links, execution sequencing, information visibility, and intermediate validation pathways while strictly maintaining the task interface and token budget.

Empirical Lead Across Complex Tasks

This runtime adaptability translates into measurable baseline improvements. Researchers evaluated MANTA across five distinct domains: information seeking, tool use, planning, workflow execution, and mathematical reasoning.

The system secured the highest average benchmark score of 74.0, outperforming the strongest baseline by 5.8 percentage points. It also captured the top score on the PlanCraft benchmark.

The Strategic Shift for Agent Builders

For engineering teams and investors, dynamic topologies change how agentic workflows scale. Rather than spending capital pre-designing rigid collaboration graphs, teams using Multi-Agent Network Topology Adaptation allow agent architectures to reconfigure themselves as task complexity changes.

System design no longer stops at deployment when runtime execution itself can reshape the underlying network architecture.