LangGraph for Go, plus the kitchen sink

What it does

tRPC-Agent-Go is a Go framework for building LLM agents with enough built-ins to fill a conference talk. It wraps models, tools, memory, and multi-agent orchestration into chainable, parallel, or graph-based runners. The project explicitly bills its GraphAgent as “functionally equivalent to LangGraph for Go.”

The interesting bit

The framework doesn’t just do agents—it does protocols. It ships with A2A (agent-to-agent), AG-UI (agent-user interaction), and MCP (Model Context Protocol) support, plus OpenTelemetry tracing and Langfuse integration. The “Skills” system is a neat touch: agents can discover and execute SKILL.md workflows from local folders or remote zip/tar archives, with caching and hot-reload via repo.Refresh().

Key highlights

• Type-safe graph workflows with conditional routing (GraphAgent)
• Chain and parallel multi-agent runners out of the box
• SKILL.md-based reusable workflows with remote loading and local caching
• Built-in evaluation framework with benchmark sets and pass/fail status
• Prompt caching claimed at “90% savings” on cached content
• Dynamic per-request agent factory for multi-tenant or sandboxed scenarios

Caveats

• The README’s “90% savings” claim for prompt caching lacks supporting data or methodology
• Several sections (Demo GIF placeholder, truncated dynamic agent example) appear unfinished
• Heavy overlap with existing Python ecosystems; the value proposition rests on “Go-native” rather than novel capabilities

Verdict Worth a look if you’re committed to Go for agent infrastructure and want protocols, observability, and evaluation in one box. Skip it if you’re already productive in Python’s agent stack and don’t need the type safety or deployment model.