// the find
The-Pocket/PocketFlow
Pocket Flow: 100-line LLM framework. Let Agents build Agents!
PocketFlow is a deliberately minimal Python LLM orchestration framework — the entire core is 100 lines built around a graph abstraction (nodes + actions as edges). It targets developers who want to wire up agent workflows without importing half of PyPI, and people using AI coding assistants to build LLM apps fast.
The 100-line core is not a gimmick — it actually covers the meaningful primitives (sync/async nodes, batch processing, flow composition) without forcing you into vendor-specific abstractions. The cookbook is genuinely useful: 35+ examples covering RAG, multi-agent, streaming, HITL, and MCP, each small enough to read in one sitting. The graph-as-first-class-concept design means complex flows like supervisor patterns and parallel batch processing are natural to express, not hacked on top of a chain abstraction. Having TypeScript, Go, Rust, Java, and C++ ports means the mental model transfers if you need to reimplement in another language.
Zero built-in observability — no tracing, no token tracking, no retry hooks baked in. You wire all of that yourself, which defeats the stated simplicity advantage when you hit production. The cookbook examples each define their own `call_llm` utility function independently, which signals that shared tooling is an afterthought rather than a design goal. State management between nodes is just a shared dict (`shared`), which scales poorly on complex flows — no type safety, no schema validation, collisions are silent. The 'agentic coding' framing (let Cursor write the code for you) is marketing for a workflow, not a feature of the framework.