Orkes
A lightweight graph framework for LLM agents. No abstractions, no black boxes, just your logic.
Orkes is a Python library for building, coordinating, and observing any workflow that can be represented as a graph. It is well suited for LLM-powered agentic systems, with a focus on explicit control flow, transparent logic, and full traceability. Its graph API is inspired by NetworkX and has no third-party dependencies.
pip install orkes
🔹 Core Concepts
- OrkesGraph: define a workflow as nodes and edges, where each node is a plain Python function
- Stateful execution: a shared state object flows between nodes during a run
- Traceability: every run produces a detailed execution trace you can inspect and visualize
🔹 Features
- Control flow: conditional branching, loops, parallel branches, and async runs with
arun - LLM integrations: built-in support for OpenAI, Anthropic Claude, and Google Gemini, including tool calls and streaming
- Agents and tools: define custom tools and use them inside graph nodes
- Live observability: send traces to Orkes Watcher with a
WatcherSink, without changing your graph code
🔹 Benchmarks
Compared with LangGraph on graphs with identical nodes and edges, Orkes runs up to 5x faster with up to 2x less memory on a 5-node graph, and up to 10x faster with up to 3x less memory on a 10-node graph.