Back to AI Research

AI Research

MANTA: Multi-Agent Network Topology Adaptation for... | AI Research

Key Takeaways

  • MANTA (Multi-Agent Network Topology Adaptation) is a framework designed to allow multi-agent systems to modify their own communication structures during the...
  • Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation.
  • However, existing systems typically treat communication topology as a fixed design choice or an offline optimization target.
  • We introduce MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time.
  • Before execution, MANTA initializes a task-conditioned topology from prior structural experience.
Paper AbstractExpand

Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation. However, existing systems typically treat communication topology as a fixed design choice or an offline optimization target. We introduce MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time. Before execution, MANTA initializes a task-conditioned topology from prior structural experience. During deployment, it monitors collaboration traces and applies bounded structural updates when the current organization becomes insufficient. These updates can modify agent roles, communication links, execution order, information visibility, and validation pathways while preserving the task interface and agent budget. We evaluate MANTA against representative single-agent and multi-agent baselines on five benchmarks spanning information seeking, tool use, planning, workflow execution, and mathematical reasoning. MANTA achieves the highest average score of 74.0, outperforming the strongest baseline by 5.8 percentage points and obtaining the best result on PlanCraft. These results show that inference-time self-improvement can extend to the architecture of collaboration itself.

MANTA (Multi-Agent Network Topology Adaptation) is a framework designed to allow multi-agent systems to modify their own communication structures during the execution of a task. While traditional multi-agent systems rely on fixed communication topologies established before deployment, MANTA enables agents to reorganize their roles, communication links, and validation pathways in response to failures or inefficiencies detected during the problem-solving process.

How MANTA Works

The framework operates through an orchestration layer that sits above the agents themselves. This layer consists of three primary components: a Topology Planner, a Trace Auditor, and a Skill Reflector.

  1. Planning: Before a task begins, the Topology Planner uses prior experience to design an initial team structure, including agent roles, interaction patterns, and information visibility policies. 2. Auditing: During execution, the Trace Auditor monitors the collaboration for process-related anomalies, such as overloaded branches or unvalidated tool failures. 3. Repair: If the current organization is insufficient, the Controller triggers a bounded structural mutation. This allows the system to add agents, rewire communication edges, or adjust information flow without updating the underlying model weights. 4. Memory: MANTA uses two "playbooks." A short-term playbook tracks the current run's history to inform immediate repairs, while a long-term playbook stores lessons across different tasks to improve future topology selection.

Performance Results

The researchers evaluated MANTA against single-agent methods, static multi-agent workflows, and automated system design methods across five benchmarks: BrowseComp, StableToolBench, PlanCraft, WorkBench, and MATH.
MANTA achieved an average score of 74.0, outperforming the strongest baseline by 5.8 percentage points. It recorded the best results on the PlanCraft benchmark. According to the researchers, the system’s ability to adapt its structure to specific task needs—rather than relying on a static design—contributes to its performance. Ablation studies indicated that the initial task-conditioned planning is the primary driver of these gains, with repair-time mutations and long-term experience providing additional improvements.

Structural Self-Improvement

The authors categorize MANTA as a "Level 6" self-improvement method, focusing on the adaptation of agent topology. Unlike methods that optimize prompts, reasoning traces, or individual agent capabilities, MANTA treats the collective organization of the team as an object of improvement. By using process-observable signals—rather than benchmark answers—to trigger repairs, the system can refine its collaboration structure in real-time.

Considerations

The framework is designed to operate without weight-level updates, relying instead on inference-time adaptation. The researchers note that while MANTA improves performance, its effectiveness depends on the ability of the Trace Auditor to identify process-level risks. The system is constrained by a fixed agent budget and requires validation of all structural changes to ensure the proposed topology remains functional. The long-term playbook is updated periodically by the Skill Reflector, which synthesizes lessons from recent execution traces to guide future task-solving.

Comments (0)

No comments yet

Be the first to share your thoughts!