Go Micro: An Agent Harness and Service Framework for Go

Go Micro is a new open-source framework from the creator of the original Go Micro project that reframes agent development as a service engineering problem. Instead of a lightweight wrapper around an LLM call, Go Micro provides a full harness: the runtime scaffolding that gives an agent model integration, persistent memory, tool access, planning, guardrails, delegation, and service discovery — all as Go code.

The core idea is that an agent is a distributed system. It needs to call real tools, preserve state across restarts, delegate sub-tasks to specialist agents, recover from failures, and be reachable over standard protocols. Go Micro's answer is to make the harness identical to the service infrastructure you already deploy. You write a service as a struct with methods; doc comments and @example tags automatically become tool descriptions for AI agents. Run it, and it's available over REST, gRPC, MCP, and an agent playground.

Agents themselves are just services with an LLM inside. They register in the service registry, maintain conversation memory in a pluggable store (Postgres, NATS KV, or file), and discover available tools from other registered services. Multiple agents coordinate via RPC — each exposes an Agent.Chat endpoint, and micro chat routes to the right one. Two built-in harness capabilities come standard: plan-delegate (an agent decomposes a goal into sub-tasks and delegates them to other agents) and dispatch (a deterministic workflow step that calls an external A2A agent). No extra graph runtime needed.

Go Micro also embraces cross-framework interoperability out of the box. It natively supports the Agent2Agent (A2A) protocol and MCP, so your agents can talk to agents built with other frameworks, and your services can expose tools for any MCP client. An A2A gateway automatically discovers agents from the registry and generates Agent Cards — no per-agent configuration. For monetization, Go Micro supports the x402 HTTP 402 payment standard, enabling tool providers to require stablecoin payments that agents can settle autonomously.

The framework comes with a CLI for scaffolding services, running them, and inspecting agents. A micro loop command even scaffolds an autonomous improvement loop — a North Star, ranked issue queue, role prompts, CI workflows, and verification gates — to let the AI improve the repository itself.

Under the hood, Go Micro is built on pluggable abstractions (registry, broker, store, model, memory, tools) with sane defaults. Memory is durable and store-backed by default, with an optional AgentCompactMemory that summaries older turns while keeping recent ones verbatim. Tools are your services automatic, plus any function registered with AgentTool.

Go Micro is still early-stage but already functional. The maintainer offers paid support and consulting. For developers building agents that must operate real systems — not just answer prompts — Go Micro provides a production-oriented foundation. ## Installation

# Install the CLI
curl -fsSL https://install.micro.mu | bash

Scaffold a service, run it, call it:

micro new service helloworld
cd helloworld
micro run

In another terminal:

micro call helloworld Helloworld.Call --name=World

Examples

• First agent on-ramp
• Examples index
• Plan-delegate example