Inherit-MAS: Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance

Inherit-MAS. Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance. By Songtao Wei and three others, from the University of Texas at Dallas and Independent Researcher. Published on the first of October. Its code repository is worked out: about one hundred and forty source files, with a readme and a licence.

Multi-agent systems built from large language models coordinate specialized agents on complex tasks, but a good workflow for them is hard to design in advance. Test-time evolution is one response: run the workflow, look at the execution feedback, refine. That loop has two costs. Broad revisions can disturb components that were already working. And re-executing requests that did not change spends computation on answers the system already had.

Inherit-MAS takes its cue from biological evolution, where inheritance and selection work together. It makes inheritance explicit at two levels. At the workflow level, a new candidate keeps what its parent got right. At the execution level, a new candidate reuses results its parent already computed.

The pieces are these. A meta-model first synthesizes a workflow of worker agents, each with a declared role, communication inputs, and tool permissions. A separately prompted judge scores each executed candidate and diagnoses what is deficient about it. In an ordinary refinement round, workflow inheritance starts from the latest completed candidate. It may discard removable nodes the judge found unhelpful, and it applies one validated edit aimed at the diagnosed deficiency, rather than rewriting the whole thing.

Execution inheritance then governs what happens when the new candidate runs. Stored results are reused only if they are eligible, and only if the complete resolved request and the execution context both match. When they do, the system skips the redundant model and tool calls.

With GPT-4o-mini workers, Inherit-MAS reaches fifty-five point four percent completion on WorkBench, and forty-nine point seven percent joint F1 on HotpotQA FullWiki. Both beat EvoAgent, EvoMAS and TacoMAS. With Qwen3-32B workers it again exceeds those evolving baselines on both benchmarks.

The savings are measured against the same controller rerun with execution inheritance switched off. Worker-token usage falls by twenty-nine point one percent on WorkBench and by thirty-four point six percent on HotpotQA. Total token usage falls by five point three percent on WorkBench, and by eighteen point one percent on HotpotQA.

The design rests on one discipline: change only what the diagnosis points at, and rerun only what actually changed. Workflow inheritance protects the parts that work from broad revision. Execution inheritance protects the budget from repeated calls. The two halves answer the two costs of test-time evolution named at the start, one each.