代理系統
MANTA Reconfigures Multi-Agent Topologies Mid-Inference, Improving the Five-Benchmark Average by 5.8 Points
MANTA builds an agent communication graph for each task, then detects collaboration anomalies from execution traces and makes constrained changes to roles, connections, and verification order. Tested with Gemma 4 31B, it achieved an average score of 74.0 across five benchmarks, although the number of samples and repeated runs remained limited.
Multi-agent systems typically fix their roles and communication workflows before deployment. The newly proposed MANTA instead treats the topology itself as a control variable at inference time. Its Topology Planner first uses the problem and a cross-task playbook to select the number of agents, their roles and groups, and the scope of information visible to them, then runs one round of collaboration. The Trace Auditor does not directly determine whether an answer is correct. Instead, it checks for observable anomalies such as tool failures, unverified evidence, premature consensus, overloaded branches, or duplicated operations. If the risk is sufficiently high, the controller permits one constrained reconfiguration, such as adding a verifier, changing connections, adjusting the execution order, or expanding a branch.
The engineering focus of the design is deterministic topology validation, agent-budget limits, deduplication of tool calls, and evidence preservation in an append-only ledger. The long-term playbook reads only process signals and has no access to benchmark answers. In theory, this allows the system to accumulate organizational experience across tasks without updating model weights.
The researchers evaluated Gemma 4 31B on 30 sampled questions each from BrowseComp, StableToolBench, PlanCraft, WorkBench, and MATH, repeating the experiments three times. MANTA achieved an average score of 74.0, 5.8 points above the runner-up, ADAS, at 68.2. On WorkBench, however, it scored only 43.3, below ADAS at 66.7. In ablation experiments, the full system achieved an average success rate of 71.7% across four benchmarks. Removing initial topology planning reduced it to 57.5%, while disabling mid-execution reconfiguration lowered it to 60.8%, indicating that the main gains did not come solely from adding more agents.
The limitations are equally clear. A full MANTA run used roughly 100,000 tokens on average, so being described as the “most economical among multi-agent systems” does not mean it is inexpensive to deploy. The evaluation set was small and used only one foundation model. The paper also did not provide an official code repository. Future work should examine whether its topology decisions remain reproducible across different models, real-world tool failures, and longer workflows, as well as whether false positives from the process auditor could damage an otherwise viable collaboration structure.