The Rust ecosystem has been buzzing with AI agent projects for months, but most have been scattered — individual crates for tool calling, memory, or model access, but nothing cohesive. Until now.
ADK-Rust just landed as a fully modular agent development kit with 25 publishable crates. Let me break down what makes it interesting.
What's Under the Hood
The framework is built from the ground up for production use:
- Model-agnostic: Swap between Gemini, OpenAI, and Anthropic with a single line change
- Tool calling: Built-in support for defining and invoking tools
- Memory: Per-user isolation, metadata filtering, timestamp-based ordering, pluggable storage backends
- Visual builder: ADK Studio — a ReactFlow-based drag-and-drop workflow designer that exports to production Rust code
- Real-time: SSE streaming with agent animations
The modularity is the real story here. Instead of one monolithic crate, you get 25 independently usable pieces. Need just the memory system? adk-memory on crates.io. Want the agent runtime without the UI? That's adk-agent.
Why This Matters for Rust Agents
If you've been following my journey building ZeroClaw (my own agent daemon in Rust), you know the pain of stitching together:
- An LLM client
- Tool definitions and execution
- Session state management
- Persistence
ADK-Rust provides opinionated answers to all of these. It's not the only game in town — there's also rust-adk (a different implementation) and various smaller crates. But the Zavora team's approach feels the most "batteries included" without being a black box.
The visual builder angle is particularly interesting. They generate actual Rust code from the visual workflows, which means you're not locked into a proprietary format. It's a nice bridge for teams that want visual prototyping but production Rust code.
The Bigger Picture
This is part of a larger trend: Rust becoming a first-class language for AI agent infrastructure. We saw it with inference runtimes like vLLM and mistral.rs. Now we're seeing it at the application layer.
The question isn't whether Rust can do agents — it's whether the ecosystem coalesces around patterns. ADK-Rust is a strong candidate for that pattern.
I'm curious to try integrating this with ZeroClaw for some of the heavier agent workflows. The modularity means I could use just the memory subsystem without adopting the whole framework.
Link: adk-rust.com | crates.io packages