a-agmon/rs-graph-llm

What it solves

graph-flow provides a way to build complex, stateful AI agent workflows in Rust. It solves the difficulty of managing long-running agent processes that need to pause for human input, branch based on LLM decisions, or resume after a system restart, all while maintaining type safety and high performance.

How it works

The framework uses a graph-based execution engine where workflows are composed of Tasks (nodes) and Edges (connections).

• State Management: A thread-safe Context stores typed key-value data and chat history, which is serialized and persisted in a Session.
• Execution Control: Tasks return a NextAction to determine the flow. This allows for "step-by-step" execution (returning control to the caller after one task) or "fire-and-forget" execution (running multiple tasks in a chain until a pause is hit).
• Routing: The engine supports standard edges, conditional edges (branching based on context data), and FanOutTask for running multiple child tasks concurrently.
• Persistence: It includes storage backends (In-memory and PostgreSQL) to ensure session state survives across process restarts.
• LLM Integration: Through an optional integration with the Rig library, it simplifies managing chat history and interacting with LLM providers.

Who it’s for

Developers building production-grade AI agents in Rust who need a structured, resumable workflow engine similar to LangGraph but with Rust's performance and type guarantees.

Highlights

• Human-in-the-loop: Built-in WaitForInput action to park workflows until external input is provided.
• Flexible Execution: Mixes interactive step-by-step execution with automated continuous chains.
• State Persistence: Native PostgreSQL support for session storage and optimistic locking to prevent race conditions.
• Parallelism: FanOutTask allows for concurrent execution of sub-tasks within a single node.
• Type-Safe Context: A serialized, typed state store that works across async tasks and synchronous edge conditions.