I build AI agents in Rust. Not because it's trendy, but because agents that run 104 tasks in parallel need zero-GC, and tools that process video/audio on device need real performance. Python for glue and ML pipelines, Rust for everything that ships.
Founder of SuperDuperAI | Blog: rustman.org
Most agent frameworks are Python wrappers around API calls. That works until you need parallel tool execution, sub-second response times, or WASM deployment. I build the infrastructure layer in Rust: the LLM client, the agent loop, the tools. Then concrete agents inherit the stack and add domain logic.
openai-oxide LLM client (caching, WebSockets, structured outputs)
└─ sgr-agent Agent framework (structured CoT, function calling, providers)
├─ sgr-agent-core Tool trait, FileBackend trait, AgentContext
├─ sgr-agent-tools 14 reusable tools (read, write, search, eval, apply_patch...)
├─ sgr-agent-ml ONNX embeddings, centroid classifier, adaptive k-NN
└─ agents built on top:
├─ agent-bit Competition agent (PAC1 benchmark, 74/104)
├─ rust-code Terminal coding agent (TUI, MCP, skills)
└─ supervox-agent Voice agent (live translate, post-call analysis)
openai-oxide is the foundation. Persistent WebSockets, SIMD JSON, hedged requests. Published on crates.io, npm, PyPI.
sgr-agent sits on top. Two-phase function calling (reasoning then action), provider routing, parallel tool execution. The FileBackend trait means the same tools work over RPC, local filesystem, or in-memory mocks.
sgr-agent-tools is the reusable toolkit: smart search (fuzzy + Levenshtein), batch read, JS eval via Boa engine, Codex-compatible diffs.
sgr-agent-ml handles on-device ML: ONNX bi-encoder embeddings, cosine-similarity classifier, adaptive k-NN with persistence. Powers agent-bit's security detection (MiniLM + DeBERTa NLI) without API calls.
| Agent | What | Score |
|---|---|---|
| agent-bit | PAC1 competition: CRM workspace, security detection, 15 skills, ONNX classifiers | 74/104 (GPT-5.4) |
| rust-code | Terminal coding agent: TUI, tmux tasks, MCP, fuzzy search | daily driver |
Both share the same stack. agent-bit adds ONNX security classifiers and a pipeline state machine. rust-code adds TUI and filesystem integration. The framework handles the common 80%.
More about the architecture: How I spent $250+ on an AI agent competition
| Project | What | Install |
|---|---|---|
| airq | Air quality CLI. Sensor + model merge, WASM core | brew install fortunto2/tap/airq |
| visa-photo | Biometric visa photos. AI background removal, Dioxus desktop | brew install fortunto2/tap/visa-photo |
| supervox | Voice toolkit. STT, VAD, TTS, mic capture | cargo add voxkit |
| OpenWok | Open-source food delivery. Dioxus fullstack, QR payments, privacy-first |
| Project | What |
|---|---|
| solo-factory | Claude Code plugin. 27 skills, 3 agents, full startup pipeline |
| solograph | Code intelligence MCP server. FalkorDB + tree-sitter, KB search, session history |
| seo-cli | SEO CLI. Google Search Console, Bing, Yandex, IndexNow |
| invoice-pdf-crm | File-based CRM. PDF invoices, letters, company cards |
Rust (agents, tools, WASM) | Python (ML, MCP servers, CLI) | TypeScript (web) | Swift (iOS) | Kotlin (Android)