PocketFlow — Minimalist 100-Line LLM Agent Framework

Introduction

PocketFlow is an ultra-minimalist LLM framework whose core fits in approximately 100 lines of Python. Despite its small size, it supports agentic workflows, RAG pipelines, multi-agent collaboration, and tool calling through a composable graph-based design.

What PocketFlow Does

• Defines workflows as directed graphs of Node objects with customizable prep, exec, and post steps
• Supports branching, looping, and conditional routing between nodes via action-based transitions
• Enables multi-agent patterns by composing multiple flows together
• Provides batch processing for parallel workload execution
• Keeps the core dependency-free so you bring your own LLM provider and tools

Architecture Overview

PocketFlow models every workflow as a graph where each Node runs three phases: prep (gather inputs), exec (call an LLM or tool), and post (process results and decide the next action). Nodes return action strings that determine graph traversal. A Flow wraps the graph and manages state as a shared dictionary passed between nodes. This architecture handles linear chains, branching decisions, retry loops, and fan-out patterns with the same primitives.

Self-Hosting & Configuration

• Install with pip install pocketflow on Python 3.8+
• No external dependencies in the core library — add your preferred LLM SDK separately
• Define nodes by subclassing Node and implementing prep, exec, and post methods
• Connect nodes with node.add_successor(next_node, action="action_name")
• Use BatchNode and BatchFlow for parallel processing of multiple items

Key Features

• Entire framework core is approximately 100 lines — easy to read, audit, and modify
• Zero vendor lock-in: works with any LLM API (OpenAI, Anthropic, local models)
• Graph-based composition supports chains, branches, loops, and nested flows
• Batch processing mode for parallelizing work across multiple inputs
• Communication patterns for multi-agent setups via shared state and nested flows

Comparison with Similar Tools

• LangChain — Feature-rich but complex; PocketFlow trades breadth for simplicity and transparency
• LangGraph — Graph-based agent framework; PocketFlow achieves similar patterns in far less code
• CrewAI — Multi-agent focus with role-based abstractions; PocketFlow is lower-level and more flexible
• DSPy — Programmatic prompt optimization; PocketFlow focuses on workflow composition
• Haystack — Pipeline-based NLP framework; PocketFlow is intentionally minimal with no built-in components

FAQ

Q: How can 100 lines of code support complex workflows? A: PocketFlow provides only the graph execution primitives. LLM calls, tool use, and retrieval are implemented in your Node subclasses using any library you choose.

Q: Does it support streaming responses? A: The core does not handle streaming directly, but your Node's exec method can use streaming LLM clients and yield results through shared state.

Q: Is PocketFlow production-ready? A: The framework is stable and used in production by multiple teams. Its simplicity makes it easy to debug and extend.

Q: Can I use it for RAG pipelines? A: Yes. Create nodes for retrieval, context assembly, and generation, then wire them into a Flow.

Sources

• https://github.com/The-Pocket/PocketFlow
• https://the-pocket.github.io/PocketFlow/