MANTA enables dynamic topology adaptation for multi-agent systems
MANTA (Multi-Agent Network Topology Adaptation) is a research framework that allows multi-agent LLM systems to dynamically reconfigure their communication topologies at inference time. By combining trace auditing with verbal playbooks during execution, it enables agent teams to optimize collaboration efficiency and achieve superior results on complex benchmarks such as PlanCraft.
Static communication hierarchies in multi-agent systems are a major bottleneck for complex reasoning, making dynamic, inference-time topology adaptation the next crucial evolutionary step. Real-time network reorganization allows agent teams to prune redundant links and establish new communication channels based on immediate execution feedback. Utilizing trace auditing alongside verbal playbooks offers an interpretable framework for agents to self-diagnose workflow failures and revise routing strategies on the fly, shifting fixed architectures toward self-evolving topological networks.
DISCOVERED
1d ago
2026-08-01
PUBLISHED
1d ago
2026-08-01
RELEVANCE
AUTHOR
Discover AI