graphbit: Build Type-Safe Multi-Agent AI Workflows

Overview

GraphBit provides a way to define AI workflows as connected nodes, including agents and tools, then execute them through a Python API backed by a Rust engine. It targets the overhead and operational challenges of coordinating multi-agent workflows, with support for concurrency, shared workflow state, and error recovery.

It may fit teams moving beyond prototypes who want a typed, observable workflow system and access to multiple LLM providers. Its performance and reliability advantages are claims made by the project; the README describes them as results from internal benchmarks.

Key Features

• Define workflows as graphs of connected agent, tool, and custom nodes.
• Use a Python API backed by Rust, with PyO3 bindings and async support.
• Connect to a range of LLM providers, including OpenAI, Anthropic, Azure OpenAI, Ollama, and OpenRouter.
• Register Python functions as tools for agents to select and call.
• Configure concurrency, retries, circuit breakers, and error handling.
• Share state across workflow steps and inspect execution results and metadata.
• Trace LLM calls and workflows, including prompts, responses, token usage, latency, and errors.

Use Cases

• Enterprise AI teams can coordinate multi-step agent workflows that need execution controls and tracing.
• Python developers can combine LLM agents and application functions in a workflow without implementing the orchestration engine themselves.
• Teams integrating several model providers can build workflows using different supported LLM services.
• Resource-conscious deployments can evaluate the Rust-backed runtime when CPU or memory overhead matters, then verify performance against their own workload.

Project Facts

• Language: Rust
• License: Apache-2.0
• Stars: 585
• Forks: 120
• Topics: agentic-ai, agentic-framework, agentic-workflow, ai, ai-agents, llm, multi-agent-systems, python, rust
• Archived: No

Getting Started

Install the Python package:

pip install graphbit

See the README for API keys, workflow examples, and further setup details.

Alternatives

• adk-rust: ADK-Rust is Rust-native and offers composable crates plus real-time voice, while GraphBit pairs a Rust core with a Python API.
• OxyGent: OxyGent is Python-first and centers on modular components and observable agent collaboration, rather than a Rust-core runtime.
• harness-sdk: Strands offers Python and TypeScript agent tools, including a ready-made harness, rather than GraphBit’s Rust-core workflow framework.
• fast-agent: fast-agent emphasizes a Python CLI, MCP server connections, and agent evaluation, while GraphBit uses a Rust core behind its Python API.

Considerations

• Workflows that call hosted LLMs require credentials for the chosen provider. The README recommends keeping API keys in environment variables or a secure secret manager.
• The framework combines a Rust runtime with a Python-facing API, so teams should be comfortable working across both layers if they need to investigate runtime or integration issues.
• Performance and determinism claims are based on the project's internal comparisons. Benchmark the framework with your own models, tools, and workload before relying on those results.
• The repository is active and licensed under Apache-2.0. Check the project documentation for current setup and API details.