Go toolkit for AI agents that won’t swallow your architecture

What it does

This is a stack of Go modules for building LLM-driven applications. The sdk layer supplies core primitives—agents, a DAG executor, telemetry—while memory adds long-term recall and hybrid retrieval, voice handles real-time STT-to-TTS pipelines over WebRTC, and vesseld offers a declarative daemon for hosting multiple agent runtimes. You can import just the pieces you need or deploy the whole thing as a single static binary.

The interesting bit

The architecture enforces a strict leaf-package boundary for sdk/engine, so new execution engines plug in by implementing a narrow Host interface rather than inheriting framework baggage. Multi-agent collaboration happens through sdk/kanban, where agents expose themselves as tools to one another using plain function calls instead of a proprietary graph DSL. Memory retrieval is deliberately hybrid: it combines BM25 with vector search and predicate-alias normalization, which suggests the authors have actually met production search quality.

Key highlights

• Provider portability across OpenAI, Anthropic, DeepSeek, MiniMax, and Volcengine via sdkx; switching providers is a YAML field change.
• vesseld runs as a standalone static binary with an HTTP + SSE control plane, Unix socket or TCP access, and bearer-token auth.
• Real-time voice pipeline with VAD, barge-in, endpointing, and WebRTC, wired to pluggable STT/TTS backends.
• Long-term memory uses hybrid BM25 + vector retrieval with predicate-alias normalization, not just embedding similarity.
• Kanban-style multi-agent delegation: any agent can be called as a tool by another agent without graph orchestration boilerplate.

Verdict

Teams already shipping Go services who need agent orchestration without importing a Python runtime or surrendering their architecture to an opinionated framework should look here. If you are looking for a no-code drag-and-drop agent builder, this is the wrong address.