agent-sdk-go

name	agent-sdk-go
description	Build production AI agents in Go with Agent SDK for Go. Use when configuring agents, tools, MCP, A2A, Temporal runtimes, streaming, memory, RAG, approvals, observability, or running examples from this documentation site.
version	1.0
compatibility	Go 1.26+. LLM API key required (OpenAI, Anthropic, or Gemini). Temporal optional for durable execution.

Agent SDK for Go

Go library for building production AI agents — LLM calls, tools, multi-turn conversation, memory, RAG, MCP and A2A integrations, human-in-the-loop approvals, and sub-agent delegation.

Full documentation index: llms.txt

Capabilities

• Create and configure agents with NewAgent and functional options
• Run agents with Run, RunAsync, or Stream
• Register built-in, custom, MCP, and A2A tools
• Persist conversation history (in-memory or Redis)
• Store and recall long-term memory (Weaviate, pgvector)
• Add RAG retrievers (agentic, prefetch, hybrid modes)
• Require human approval for tools and sub-agent delegation
• Attach middleware hooks at LLM, tool, retrieval, and memory lifecycle points
• Export OpenTelemetry traces, metrics, and logs
• Execute in-process or on Temporal for durable workflows

Workflows

Create a minimal agent

1. Read Quickstart
2. Configure an LLM client — LLM Providers
3. Call NewAgent with WithLLMClient and WithSystemPrompt
4. Call Run(ctx, prompt, nil) and read AgentRunResult
5. Always call defer a.Close() — required to flush OTLP exporters on shutdown

Example: Simple Agent

Add tools

1. Read Tools
2. Register tools with WithTools or WithToolRegistry
3. Set WithToolApprovalPolicy(AutoToolApprovalPolicy()) for trusted automation
4. Run: Tools example

Stream to a UI

1. Read Streaming
2. Pass WithStream(true) at agent creation
3. Call Stream and consume <-chan AgentEvent
4. Check for nil events; handle RUN_FINISHED for final result and token usage
5. Example: Stream · AG-UI

Switch to Temporal (durable execution)

1. Read Temporal runtime
2. Add WithTemporalConfig or WithTemporalClient — never both
3. For production, split client and worker — Worker separation
4. Align agent and worker configuration (fingerprint) — same name, LLM, tools, hooks group names, approval policy
5. Examples: Temporal Client · Agent Worker · Durable Agent

Integration

Component	Documentation
LLM providers	LLM Providers
All agent options	Configuration
In-process runtime	In-Process
Temporal runtime	Temporal
MCP servers	MCP
A2A remote agents	A2A
Observability	Telemetry
Runnable examples	Running Examples
Go API reference	https://pkg.go.dev/github.com/agenticenv/agent-sdk-go

Context

• Architecture: Architecture
• Agent loop: prepare context → call LLM → execute tools (iterate) → finalize
• Capabilities resolve at call time from registries — tools, MCP, A2A, and sub-agents can change between runs
• Feature pages explain concepts; example pages show run commands and expected output under examples/
• Default tool approval policy is require-all — set AutoToolApprovalPolicy() for unattended runs
• With DisableLocalWorker() and streaming, you must also call EnableRemoteWorkers()
• Hook group names participate in the Temporal agent fingerprint — register the same names on client and worker

Documentation map

Section	Entry point
Overview	Introduction
Getting started	Quickstart
Features	Tools
Advanced	Worker separation
Observability	Telemetry
Examples	Running Examples
Production	Readiness