When developers build LLM‑powered agents, they often wrestle with heavyweight frameworks that add graphs, nodes, and planners. PicoFlow sidesteps that by treating a workflow as a chain of normal async Python functions. The result is a tiny, async‑first DSL that keeps logic in the code, not in opaque graph layers.

The library exposes a simple flow decorator and an llm helper that accepts a prompt template and a URL‑style adapter. Switching providers becomes a matter of changing the URL, leaving the workflow untouched. Debugging stays local because stack traces never jump across hidden schedulers or internal layers for reproducibility and traceability in production today.

For small to medium projects—CLI agents, backend pipelines, or RAG prototypes—PicoFlow offers readability, composability, and low cognitive overhead. It isn’t a full agent OS, but it fills the gap left by LangChain and CrewAI. Open‑source contributors can tweak the DSL, add tracing hooks, or build custom control‑flow helpers.