Multi-agent systems evolve workflows to reduce redundant computation
Inherit-MAS: Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance
Machine LearningArtificial Intelligence
Summary
Multi-agent systems made from language models help solve complex tasks by dividing work among specialized agents, but designing good workflows beforehand is hard. The authors propose Inherit-MAS, a method that evolves these workflows during test time by learning from past execution feedback and reusing useful parts without repeating work unnecessarily. It borrows ideas from biological inheritance to make workflow changes more efficient and reduces the amount of repeated computation. Their approach outperforms earlier systems on complex task benchmarks and uses fewer computational resources.
What this means in practice
- •For software engineering teams: Improve complex system automation by evolving and refining agent workflows during operation to reduce resource waste and increase task completion rates.
- •For customer support centers: Enhance automated multi-agent customer handling by evolving coordination workflows to respond more effectively to varied and complex queries.
Authors
Songtao Wei, Yi Li, Zhichun Guo, Bingzhe Li
Abstract
Multi-agent systems (MAS) built from large language models coordinate specialized agents to tackle complex tasks, but effective workflows are difficult to design in advance. Test-time evolution refines workflows using execution feedback, yet broad revisions can disturb useful components, while re-executing unchanged requests can incur redundant computation. Inspired by the interplay of inheritance and selection in biological evolution, we introduce Inherit-MAS, which makes inheritance explicit at the workflow and execution levels. A meta-model first synthesizes a workflow of worker agents with declared roles, communication inputs, and tool permissions, and a separately prompted judge scores each executed candidate and diagnoses its deficiencies. In ordinary refinement rounds, \emph{workflow inheritance} starts from the latest completed candidate, may discard removable nodes judged unhelpful, and applies a validated edit to address the diagnosed deficiency. When the new candidate executes, \emph{execution inheritance} inherits eligible stored results only if the complete resolved request and execution context match, avoiding redundant model and tool calls. With GPT-4o-mini workers, Inherit-MAS achieves 55.4\% completion on WorkBench and 49.7\% joint F1 on HotpotQA FullWiki, outperforming EvoAgent, EvoMAS, and TacoMAS. With Qwen3-32B workers, it also exceeds these evolving-MAS baselines on both benchmarks. Compared with rerunning the same controller with execution inheritance disabled, execution inheritance reduces worker-token usage by 29.1\% on WorkBench and 34.6\% on HotpotQA, and total token usage by 5.3\% and 18.1\%.