Rig

Rig's pitch is narrower and more honest than most agent frameworks: it is a Rust library, and that is the whole trade. What you get in return is a type system doing work that other stacks do at runtime. Tool arguments and structured outputs are plain Rust types, so a mismatch between what your model returns and what your code expects is caught by the compiler rather than by a 3am alert. The docs frame this as "type-safe by construction," and the mechanics back it up — the same agent, completion, and embedding abstractions work across 20+ providers, so swapping OpenAI for Anthropic or Gemini is a client change, not a rewrite. The surrounding capabilities are broader than a thin chat wrapper: native RAG through embeddings and vector store indexes, multi-agent systems that coordinate specialized agents, advanced model routing, dynamic model creation, memory, loaders, media handling, observability, evals, and Model Context Protocol support. Documented vector stores include LanceDB, MongoDB, Neo4j, Qdrant, SurrealDB, and an in-memory option, with AWS Lambda deployment guides including a LanceDB-on-Lambda walkthrough. Production tooling is called out explicitly in the docs: typed errors, token-usage accounting, tracing, and mock models for offline deterministic testing. That last group is where Rig earns its keep — without mocks, LLM tests are flaky and expensive, and Rig's testing story is a first-class feature rather than an afterthought. Portability is the other differentiator: you compile to a single self-contained executable or to WebAssembly for edge and browser targets, with no runtime or dependency tree to ship. The honest weaknesses: this is Rust-only, and async/await plus the trait system are table stakes — the quickstart itself requires Tokio's macros and multi-threaded runtime features. There is no GUI, no no-code builder, and no hosted service; it is a crate you integrate. Custom providers or vector stores mean extending the library rather than installing a plugin. Ecosystem breadth is the price paid for speed and compile-time safety when compared with Python-side frameworks. On sustainability: the June 2026 post marking one million downloads, adding a maintainer, and publishing a roadmap is a genuinely positive signal for an open-source framework, since single-maintainer risk is the usual reason teams hesitate. Rig fits teams already writing async Rust who want provider-agnostic AI without a second language, and it fits edge or WASM deployments where shipping a binary or browser module is the point. It does not fit Python-first data science, prompt experimentation without a deploy target, or anyone who needs a visual agent editor.