agnt-core

The zero-I/O kernel of the agnt agent runtime.

Defines the message types, tool trait, backend abstraction, persistence abstraction, observer hooks, and the synchronous agent loop itself — with no HTTP, no SQLite, and no async runtime dependencies.

[dependencies]
agnt-core = "0.2"

When to depend on this

Most users should cargo add agnt instead — the flagship crate pulls in agnt-core plus agnt-net, agnt-store, and agnt-tools as optional features and gives you a working agent runtime.

Depend directly on agnt-core when:

• You want the minimum possible dependency footprint (only serde, serde_json, tracing)
• You're targeting WASM — agnt-core has no filesystem, TLS, or process-spawning deps that would block wasm32-* builds
• You're bringing your own backend (implementing [LlmBackend]) and your own persistence (implementing [MessageStore])
• You're embedding the agent loop into a larger system with pre-existing HTTP and persistence layers

What it provides

• [Agent<B: LlmBackend>] — the core loop (message → inference → parallel tool dispatch → loop)
• [AgentBuilder] — fluent construction
• [Tool] (erased) and [TypedTool] (typed Args/Output/Error) traits, with [ErasedAdapter] bridging typed impls into the erased dispatch path
• [Registry] — name-based tool dispatch
• [Message], [ToolCall], [FunctionCall] — the internal wire format
• [LlmBackend] and [BackendError] — trait any backend implements
• [MessageStore], [StoreError], [ToolLog] — trait any store implements
• [Observer] — lifecycle hook trait for HITL approval, audit, metrics
• tracing instrumentation at agnt.step, agnt.backend.chat, agnt.tool span boundaries

Security

See the threat model. agnt-core handles tool output envelope framing (<tool_output>) and enforces the 64KB per-result byte cap. Filesystem sandboxing and SSRF guards live in agnt-tools.

License

Dual-licensed under MIT OR Apache-2.0.