Pay for the Fault, Not the Flow: Label-Free In-Flow Multi-Agent Workflow Optimization

Pay for the Fault, Not the Flow. Label-Free In-Flow Multi-Agent Workflow Optimization. By Xuehang Guo and six others, from William and Mary and NEC Corporation of America. Published on the first of October. No code repository is linked to it.

Large language models increasingly construct multi-agent workflows. They decompose a complex task and assign specialist agents from a pool to the pieces. Building such a workflow well is hard, because several critical decisions have to be settled up front. How finely to divide the task. Which agent to trust with each subtask. When to create a new specialist instead. Whether each subtask actually succeeds stays unknown until the workflow runs. Improving a workflow afterwards is costly on two counts. Locating a fault usually requires a reference answer, a graded outcome, or a trained assessor. And the fix is applied to the whole workflow, through re-execution, re-search, or retraining.

InFlowOp puts every one of those decisions on a single price. It defines one label-free cost, which weighs how well an agent's competence meets what a subtask demands against how much that agent takes to run. The title states the principle: pay for the fault, not the flow. Two properties follow from the design. Because the cost is label-free, it does not wait on a reference answer or a trained assessor. And because it is one cost, the same quantity can serve the workflow both while it is being built and while it is running.

Before execution, InFlowOp uses the cost to settle decomposition and assignment together. It works bidirectionally, determining the granularity of task decomposition and the assignment of agents from the cost rather than from a fixed template. During execution, when a fault appears, InFlowOp corrects it with the cheapest move available, priced by that same cost. Construction and repair are therefore not two separate systems with two separate criteria. They are one criterion applied at two moments.

Evaluation needed its own answer. Judging a whole workflow, rather than a single agent, is a challenge in itself, so the authors introduce Braid, a benchmark whose tasks require multi-agent coordination beyond what a single agent is capable of.

Across various domains and backbones, InFlowOp outperforms single-agent baselines by up to eleven point nine seven percent. With in-flow optimization, the correction made while the workflow runs, the reported figure is nine point six four percent. Those numbers are set against single agents, on tasks built so that a single agent should not be enough.

Where the design concentrates its weight is a single number. Competence against demand, set against the price of running an agent, decides how a task is split, who handles each piece, and what to change when something breaks. A fault is then fixed by its cheapest move, not by paying to rerun the whole flow.