agents-go

A Go framework for building LLM agents. Uses any-llm-go as the LLM provider abstraction. Not feature complete yet.

Features

• Agent loop — model call → tool execution → model call, with parallel tool execution
• Streaming — channel-based streaming with text deltas and tool call events
• Structured output — provider-native (json_schema) or tool-based strategy, with auto-fallback
• Middleware — composable hooks for model calls, tool calls, and agent lifecycle
• Multi-agent — compose agents as tools for orchestrator/specialist patterns
• MCP — connect to MCP servers (streamable HTTP) and use their tools
• Built-in tools — task tracking (write_todos)

Install

go get github.com/rashadism/agents-go

Quick start

a, _ := agent.CreateAgent(provider, "gpt-5.2",
    agent.WithSystemPrompt("You are a helpful assistant."),
    agent.WithTools(myTool),
)

result, _ := a.Run(ctx, []providers.Message{
    {Role: providers.RoleUser, Content: "Hello"},
})

Multi-agent composition

// Create a specialist sub-agent with its own tools and system prompt
researcher, _ := agent.CreateAgent(provider, "gpt-5.2",
    agent.WithSystemPrompt("You are a research specialist."),
    agent.WithTools(webSearch),
)

// Give it to a parent agent as a tool
orchestrator, _ := agent.CreateAgent(provider, "gpt-5.2",
    agent.WithSystemPrompt("You coordinate research tasks."),
    agent.WithTools(
        agent.AgentTool("researcher", "Delegates research tasks", researcher),
    ),
)

result, _ := orchestrator.Run(ctx, messages)

Examples

go run ./examples/basic                  # Travel assistant — single tool
go run ./examples/streaming              # Travel assistant — streaming with parallel tool calls
go run ./examples/middleware              # Customer support — logging and message-budget middleware
go run ./examples/structured-output      # Recipe generator — provider-native structured output
go run ./examples/streaming-structured   # Product comparison — streaming + structured output
go run ./examples/multi-agent            # Deep research — orchestrator + researcher sub-agents (requires TAVILY_API_KEY)