PlanCraft

MANTA: Framework for self-evolving multi-agent communication topology at inference time

MANTA is a new framework that allows multi-agent LLM systems to dynamically adapt their communication topology during inference rather than treating it as a fixed design choice. The system initializes a task-conditioned topology from prior structural experience, then monitors collaboration traces and applies bounded structural updates—modifying agent roles, communication links, execution order, and validation pathways—as needed. Evaluated across five benchmarks spanning information seeking, tool use, planning, workflow execution, and mathematical reasoning, MANTA achieves an average score of 74.0, outperforming the strongest baseline by 5.8 percentage points.