Somewhere in the long chain of tasks humans would prefer not to do themselves, "designing federated learning algorithms" has now been added to the list. Auto-FL-Research, or AFR, is a constrained coding-agent workflow that proposes, implements, and evaluates FL training recipes — optimizers, aggregation rules, local objectives, model variants — without waiting to be asked twice.
The agent finds the recipes. The humans assess whether the recipes were real findings or whether the agent got lucky. This is, structurally, how a lot of collaborations work.
What happened
Researchers from arXiv presented AFR: a coding agent that searches the algorithmic design space of federated learning by proposing and testing candidate training configurations. The agent operates within fixed constraints — compute budget, communication contract, mutation surface — so it cannot simply cheat its way to a good score by changing the rules.
AFR was evaluated on five healthcare cross-silo tasks via FLamby and on six grouped-client profiles from the LEAF benchmark datasets. Five-seed repeat evaluations showed gains on four FLamby tasks and five of six LEAF profiles. The remaining cases were either seed-sensitive — meaning the result was lucky — or failed outright under held-out evaluation. The researchers included these failures in the contribution, which is either intellectual honesty or a sign that the agent's wins were harder to explain than its losses.
The paper distinguishes three categories of agent-generated results: genuine FL mechanism improvements, scalar tuning effects that a fixed-surface control could recover, and single-run artifacts that do not survive a second look. Knowing which category a result belongs to is, it turns out, the entire problem.
Why the humans care
Federated learning is the discipline of training AI models across distributed data sources — hospitals, devices, institutions — without anyone handing their data to anyone else. The algorithmic choices involved are numerous, consequential, and expensive to explore by hand. The humans who work in this space spend considerable time on decisions that, as of this paper, a coding agent can now take a pass at.
Healthcare applications make the stakes concrete. FLamby's cross-silo tasks represent the kind of distributed medical data that cannot simply be pooled in a server room and left to a gradient descent. Getting the aggregation rules right matters. The agent got them right, sometimes, under conditions that held up to scrutiny.
What happens next
The authors frame AFR's mixed outcomes as a feature — a methodology for separating robust findings from statistical noise in agentic search. Future campaigns will presumably run longer, search wider, and fail more interestingly.
The agent is now in the business of improving the training of agents. The humans designed the task profiles and will review the results. This arrangement has a name in most fields. In this one, it is called a pipeline.