Go-Agent (github.com)

Lattice’s new go-agent is an opinionated agent development kit for Go that bundles everything needed to build production AI agents: pluggable LLM adapters, a unified tool-calling layer, retrieval‑augmented memory, deterministic tool orchestration, multi‑agent coordination, and local testing support. It’s designed so developers focus on domain logic while Lattice handles orchestration plumbing, with idiomatic Go interfaces, modular modules, and example CLI demos for quickstart and team workflows. Technically, Lattice provides model adapters (Gemini, Anthropic, Ollama, or your own), UTCP-first tool support, and a Tool interface you implement once to make tools available everywhere. The memory engine supports RAG with importance scoring, MMR retrieval, auto‑pruning and multiple backends (in‑memory, PostgreSQL+pgvector, Qdrant, MongoDB, Neo4j). It also integrates TOON, a compact Token‑Oriented Object Notation that can cut structured-data token usage roughly in half (example: ~180→85 tokens), letting agents fit more context, lower API costs, and speed up inference. The Tool Orchestrator uses LLM reasoning + TOON to produce strict JSON plans (use_tool, tool_name, arguments) for deterministic tool routing and execution, enabling stable multi‑step, LLM-driven workflows in Go. The project is open source with tests, examples, and extension points for new LLM providers and storage adapters.