Agent SDK for Go

​

Your First Agent

llmClient, _ := openai.NewClient(
    llm.WithAPIKey(os.Getenv("OPENAI_API_KEY")),
    llm.WithModel("gpt-4o"),
)

a, _ := agent.NewAgent(
    agent.WithName("my-agent"),
    agent.WithSystemPrompt("You are a helpful assistant."),
    agent.WithLLMClient(llmClient),
)
defer a.Close()

agentRun, _ := a.Run(ctx, "What is the capital of France?", nil)
result, _ := agentRun.Get(ctx)
fmt.Println(result.Content) // Paris.

​

Install

go get github.com/agenticenv/agent-sdk-go@latest

​

Why Agent SDK for Go

• Idiomatic Go — functional options, typed interfaces, no reflection magic. Swap any component (LLM client, memory backend, approval policy) by passing a different option.
• Durable by default, no infrastructure — the in-process runtime journals every step via durable-go; kill the process and reconnect after a restart. Add temporal.WithTemporalConfig (from pkg/agent/runtime/temporal) or restate.WithRestateConfig (from pkg/agent/runtime/restate) only when you need horizontal scale or a client/worker split. Nothing else changes.
• Concurrent by default — one Agent instance handles parallel Run and Stream calls; every runtime issues each a unique run automatically.
• Protocol-native integrations — MCP tool servers, A2A agent-to-agent delegation, and AG-UI streaming events are first-class, not adapters bolted on after the fact.
• Middleware hooks — intercept LLM calls, tool use, retrieval, and memory at any lifecycle point for logging, PII scrubbing, and guardrails without touching agent logic.
• Error control — after retries, swap to a fallback model, grant more iterations, or skip a looping tool. See Error Control.
• Cache-aware requests — LLM requests are structured for provider caching, with Anthropic prompt-cache breakpoints to reduce cost on multi-turn runs.

​

Runtimes

​

Start here