Rust agent base that separates the kernel from the kitchen sink

What it does

Loong is a 13-crate Rust workspace acting as a controlled base for vertical AI agents. It exposes a TUI and CLI surface for longer-horizon workflows, compound tasks, and human-in-the-loop collaboration, while keeping providers, tools, memory, and policy outside the kernel so you compile in only what you need. The project also bakes in first-class integration with Lark/Feishu, suggesting its authors expect agents to live inside enterprise chat threads as much as local terminals.

The interesting bit

The architecture enforces a strict acyclic dependency graph across thirteen crates, splitting stable contracts, runtime substrate, validation rails, and delivery surfaces into governed layers. That is unusual in the agent-framework space, where most projects quickly become a flat pile of Python glue; Loong instead treats modularity as a compile-time guarantee.

Key highlights

• 42+ built-in providers and 25+ channels out of the box, including Feishu/Lark enterprise chat integration.
• Explicit runtime boundaries around provider selection, tools, memory, approvals, policy, and audit.
• Configuration compatible with existing setups from OpenClaw, Claude Code, Codex, and similar tools.
• Core and extensions are physically separate crates; you compile and compose providers, channels, and memory as needed rather than inheriting a monolith.
• Rust Edition 2024 codebase with MIT licensing.

Verdict

Worth a look if you are building production-adjacent agents in Rust and want an opinionated, layered architecture with compile-time boundaries. Skip it if you need a zero-config, batteries-included framework that does not require compiling a 13-crate workspace.