tRPC-Agent-Go

tRPC-Agent-Go is an open-source Go framework from the tRPC team (trpc-group) for building agent applications. It combines LLM agents, graph workflows, tool calling, session and memory state, knowledge retrieval, evaluation, and OpenTelemetry observability in one Go-native stack. The latest listed release is v1.11.2, and the code is licensed under Apache 2.0.

What It Is

tRPC-Agent-Go is a developer framework for Go services that need LLM-driven agents. A Runner executes an Agent against a model, tools, session state, and memory, and streams events back to the caller. The documentation says the project aims to bring autonomous multi-agent capability to Go, leveraging Go's concurrency and the tRPC ecosystem.

How the Agent System Works

The architecture is modular and pluggable, with components communicating through an event-driven mechanism and supporting callbacks for custom logic. Core components are Agent, Runner, Model (OpenAI, DeepSeek, etc.), Tool, Session, Memory, Knowledge (RAG), and Planner.

Built-in agent types:

• LLMAgent: model-based agent with tool calling and reasoning.
• ChainAgent: sequential sub-agents.
• ParallelAgent: concurrent sub-agents with merged output.
• CycleAgent: iterative loops.
• GraphAgent: type-safe graph workflows with conditional and multi-conditional routing.

Tools, Skills, and Self-Evolution

Any Go function can become a tool, and MCP tools, web search (DuckDuckGo), and code execution are supported. Agent Skills are folders with a SKILL.md spec that agents load and run on demand. An evolution service reviews completed sessions asynchronously, passes candidates through quality gates, and publishes reusable procedures as managed skills.

State, Evaluation, and Observability

Session and memory backends documented include in-memory, Redis, SQLite, MySQL, PostgreSQL, pgvector, ClickHouse and MongoDB, while knowledge retrieval supports vector stores such as pgvector, Elasticsearch, Qdrant, Milvus, and ChromaDB. Evaluation uses eval sets and metrics to measure agent quality. OpenTelemetry tracing and metrics are supported, with Langfuse examples.

Server and Protocol Integration

Servers can expose agents through AG-UI (with SSE streaming), A2A, and a gateway, and the docs include an OpenClaw runtime. The README lists use cases such as customer support bots, data analysis assistants, DevOps automation, and RAG-powered knowledge agents.