ControlFlow is built around a core opinion that LLMs perform really well on small, well-defined tasks and run off the rails otherwise. I know that may seem obvious, but the key insight is that if you compose enough of these small tasks into a structured workflow, you can recover the kind of complex behaviors we associate with autonomous AIs, without sacrificing control or observability at each step. It ends up feeling a lot like writing a traditional software workflow.

With ControlFlow you:

- define structured tasks for your agents to achieve,

- assign one or more specialized agents to each task,

- then compose the tasks into a workflow for execution.

We've been working toward this for a while, applying insights from our work on Prefect (our workflow engine) and Marvin (our LLM framework). ControlFlow is our effort to bridge the gap between all those impressive AI demos and what we see B2B users actually trying to accomplish in production. More often than not, their goal is to sanely delegate a few steps of a workflow to an AI, not deploy a suite of end-to-end autonomous while loops.

This is ControlFlow's initial release and we’d love your feedback. We hope it's a refreshing take on making LLMs feel like software again! You can check out the repo here: https://github.com/prefecthq/controlflow/