Add agent sub-package for composable LLM agents

Introduces v2/agent with a minimal API: Agent, New(), Run(), and AsTool(). Agents wrap a model + system prompt + tools. AsTool() turns an agent into a llm.Tool, enabling parent agents to delegate to sub-agents through the normal tool-call loop — no channels, pools, or orchestration needed. Also exports NewClient(provider.Provider) for custom provider integration. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-07 23:17:19 -05:00
parent be572a76f4
commit 87ec56a2be
6 changed files with 472 additions and 7 deletions
--- a/v2/agent/agent.go
+++ b/v2/agent/agent.go
@@ -0,0 +1,113 @@
+// Package agent provides a simple agent abstraction built on top of go-llm.
+//
+// An Agent wraps a model, system prompt, and tools into a reusable unit.
+// Agents can be turned into tools via AsTool, enabling parent agents to
+// delegate work to specialized sub-agents through the normal tool-call loop.
+//
+// Example — orchestrator with sub-agents:
+//
+//	researcher := agent.New(model, "You research topics via web search.",
+//	    agent.WithTools(llm.NewToolBox(tools.WebSearch(apiKey))),
+//	)
+//	coder := agent.New(model, "You write and run code.",
+//	    agent.WithTools(llm.NewToolBox(tools.Exec())),
+//	)
+//	orchestrator := agent.New(model, "You coordinate research and coding tasks.",
+//	    agent.WithTools(llm.NewToolBox(
+//	        researcher.AsTool("research", "Research a topic"),
+//	        coder.AsTool("code", "Write and run code"),
+//	    )),
+//	)
+//	result, err := orchestrator.Run(ctx, "Build a fibonacci function in Go")
+package agent
+
+import (
+	"context"
+
+	llm "gitea.stevedudenhoeffer.com/steve/go-llm/v2"
+)
+
+// Agent is a configured LLM agent with a system prompt and tools.
+// Each call to Run creates a fresh conversation (no state is carried between runs).
+type Agent struct {
+	model   *llm.Model
+	system  string
+	tools   *llm.ToolBox
+	reqOpts []llm.RequestOption
+}
+
+// Option configures an Agent.
+type Option func(*Agent)
+
+// WithTools sets the tools available to the agent.
+func WithTools(tb *llm.ToolBox) Option {
+	return func(a *Agent) { a.tools = tb }
+}
+
+// WithRequestOptions sets default request options (temperature, max tokens, etc.)
+// applied to every completion call the agent makes.
+func WithRequestOptions(opts ...llm.RequestOption) Option {
+	return func(a *Agent) { a.reqOpts = opts }
+}
+
+// New creates an agent with the given model and system prompt.
+func New(model *llm.Model, system string, opts ...Option) *Agent {
+	a := &Agent{
+		model:  model,
+		system: system,
+	}
+	for _, opt := range opts {
+		opt(a)
+	}
+	return a
+}
+
+// Run executes the agent with a user prompt. Each call is a fresh conversation.
+// The agent loops tool calls automatically until it produces a text response.
+func (a *Agent) Run(ctx context.Context, prompt string) (string, error) {
+	return a.RunMessages(ctx, []llm.Message{llm.UserMessage(prompt)})
+}
+
+// RunMessages executes the agent with full message control.
+// Each call is a fresh conversation. The agent loops tool calls automatically.
+func (a *Agent) RunMessages(ctx context.Context, messages []llm.Message) (string, error) {
+	chat := llm.NewChat(a.model, a.reqOpts...)
+	if a.system != "" {
+		chat.SetSystem(a.system)
+	}
+	if a.tools != nil {
+		chat.SetTools(a.tools)
+	}
+
+	// Send each message; the last one triggers the completion loop.
+	// All but the last are added as context.
+	for i, msg := range messages {
+		if i < len(messages)-1 {
+			chat.AddToolResults(msg) // AddToolResults just appends to history
+			continue
+		}
+		return chat.SendMessage(ctx, msg)
+	}
+
+	// Empty messages — send an empty user message
+	return chat.Send(ctx, "")
+}
+
+// delegateParams is the parameter struct for the tool created by AsTool.
+type delegateParams struct {
+	Input string `json:"input" description:"The task or question to delegate to this agent"`
+}
+
+// AsTool creates a llm.Tool that delegates to this agent.
+// When a parent agent calls this tool, it runs the agent with the provided input
+// as the prompt and returns the agent's text response as the tool result.
+//
+// This enables sub-agent patterns where a parent agent can spawn specialized
+// child agents through the normal tool-call mechanism.
+func (a *Agent) AsTool(name, description string) llm.Tool {
+	return llm.Define[delegateParams](name, description,
+		func(ctx context.Context, p delegateParams) (string, error) {
+			return a.Run(ctx, p.Input)
+		},
+	)
+}
--- a/v2/agent/agent_test.go
+++ b/v2/agent/agent_test.go
@@ -0,0 +1,244 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+
+	llm "gitea.stevedudenhoeffer.com/steve/go-llm/v2"
+	"gitea.stevedudenhoeffer.com/steve/go-llm/v2/provider"
+)
+
+// mockProvider is a test helper that implements provider.Provider.
+type mockProvider struct {
+	mu           sync.Mutex
+	completeFunc func(ctx context.Context, req provider.Request) (provider.Response, error)
+	requests     []provider.Request
+}
+
+func (m *mockProvider) Complete(ctx context.Context, req provider.Request) (provider.Response, error) {
+	m.mu.Lock()
+	m.requests = append(m.requests, req)
+	m.mu.Unlock()
+	return m.completeFunc(ctx, req)
+}
+
+func (m *mockProvider) Stream(ctx context.Context, req provider.Request, events chan<- provider.StreamEvent) error {
+	close(events)
+	return nil
+}
+
+func (m *mockProvider) lastRequest() provider.Request {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if len(m.requests) == 0 {
+		return provider.Request{}
+	}
+	return m.requests[len(m.requests)-1]
+}
+
+func newMockModel(fn func(ctx context.Context, req provider.Request) (provider.Response, error)) *llm.Model {
+	mp := &mockProvider{completeFunc: fn}
+	return llm.NewClient(mp).Model("mock-model")
+}
+
+func newSimpleMockModel(text string) *llm.Model {
+	return newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		return provider.Response{Text: text}, nil
+	})
+}
+
+func TestAgent_Run(t *testing.T) {
+	model := newSimpleMockModel("Hello from agent!")
+	a := New(model, "You are a helpful assistant.")
+
+	result, err := a.Run(context.Background(), "Say hello")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "Hello from agent!" {
+		t.Errorf("expected 'Hello from agent!', got %q", result)
+	}
+}
+
+func TestAgent_Run_WithTools(t *testing.T) {
+	callCount := 0
+	model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		callCount++
+		if callCount == 1 {
+			// First call: model requests a tool call
+			return provider.Response{
+				ToolCalls: []provider.ToolCall{
+					{ID: "tc1", Name: "greet", Arguments: `{}`},
+				},
+			}, nil
+		}
+		// Second call: model returns text after seeing tool result
+		return provider.Response{Text: "Tool said: hello!"}, nil
+	})
+
+	tool := llm.DefineSimple("greet", "Says hello", func(ctx context.Context) (string, error) {
+		return "hello!", nil
+	})
+
+	a := New(model, "You are helpful.", WithTools(llm.NewToolBox(tool)))
+	result, err := a.Run(context.Background(), "Use the greet tool")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "Tool said: hello!" {
+		t.Errorf("expected 'Tool said: hello!', got %q", result)
+	}
+	if callCount != 2 {
+		t.Errorf("expected 2 calls (tool loop), got %d", callCount)
+	}
+}
+
+func TestAgent_AsTool(t *testing.T) {
+	// Create a child agent
+	childModel := newSimpleMockModel("child result: 42")
+	child := New(childModel, "You compute things.")
+
+	// Create the tool from the child agent
+	childTool := child.AsTool("compute", "Delegate computation to child agent")
+
+	// Verify tool metadata
+	if childTool.Name != "compute" {
+		t.Errorf("expected tool name 'compute', got %q", childTool.Name)
+	}
+	if childTool.Description != "Delegate computation to child agent" {
+		t.Errorf("expected correct description, got %q", childTool.Description)
+	}
+
+	// Execute the tool directly (simulating what the parent's Chat.Send loop does)
+	result, err := childTool.Execute(context.Background(), `{"input":"what is 6*7?"}`)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "child result: 42" {
+		t.Errorf("expected 'child result: 42', got %q", result)
+	}
+}
+
+func TestAgent_AsTool_ParentChild(t *testing.T) {
+	// Child agent that always returns a fixed result
+	childModel := newSimpleMockModel("researched: Go generics are great")
+	child := New(childModel, "You are a researcher.")
+
+	// Parent agent: first call returns tool call, second returns text
+	parentCallCount := 0
+	parentModel := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		parentCallCount++
+		if parentCallCount == 1 {
+			return provider.Response{
+				ToolCalls: []provider.ToolCall{
+					{ID: "tc1", Name: "research", Arguments: `{"input":"Tell me about Go generics"}`},
+				},
+			}, nil
+		}
+		// After getting tool result, parent synthesizes final answer
+		return provider.Response{Text: "Based on research: Go generics are great"}, nil
+	})
+
+	parent := New(parentModel, "You coordinate tasks.",
+		WithTools(llm.NewToolBox(
+			child.AsTool("research", "Research a topic"),
+		)),
+	)
+
+	result, err := parent.Run(context.Background(), "Tell me about Go generics")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "Based on research: Go generics are great" {
+		t.Errorf("expected synthesized result, got %q", result)
+	}
+	if parentCallCount != 2 {
+		t.Errorf("expected 2 parent calls (tool loop), got %d", parentCallCount)
+	}
+}
+
+func TestAgent_RunMessages(t *testing.T) {
+	model := newSimpleMockModel("I see the system and user messages")
+	a := New(model, "You are helpful.")
+
+	messages := []llm.Message{
+		llm.UserMessage("First question"),
+		llm.AssistantMessage("First answer"),
+		llm.UserMessage("Follow up"),
+	}
+
+	result, err := a.RunMessages(context.Background(), messages)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "I see the system and user messages" {
+		t.Errorf("unexpected result: %q", result)
+	}
+}
+
+func TestAgent_ContextCancellation(t *testing.T) {
+	model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		return provider.Response{}, ctx.Err()
+	})
+	a := New(model, "You are helpful.")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // Cancel immediately
+
+	_, err := a.Run(ctx, "This should fail")
+	if err == nil {
+		t.Fatal("expected error from cancelled context")
+	}
+}
+
+func TestAgent_WithRequestOptions(t *testing.T) {
+	var capturedReq provider.Request
+	model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		capturedReq = req
+		return provider.Response{Text: "ok"}, nil
+	})
+
+	a := New(model, "You are helpful.",
+		WithRequestOptions(llm.WithTemperature(0.3), llm.WithMaxTokens(100)),
+	)
+
+	_, err := a.Run(context.Background(), "test")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if capturedReq.Temperature == nil || *capturedReq.Temperature != 0.3 {
+		t.Errorf("expected temperature 0.3, got %v", capturedReq.Temperature)
+	}
+	if capturedReq.MaxTokens == nil || *capturedReq.MaxTokens != 100 {
+		t.Errorf("expected maxTokens 100, got %v", capturedReq.MaxTokens)
+	}
+}
+
+func TestAgent_Run_Error(t *testing.T) {
+	wantErr := errors.New("model failed")
+	model := newMockModel(func(ctx context.Context, req provider.Request) (provider.Response, error) {
+		return provider.Response{}, wantErr
+	})
+	a := New(model, "You are helpful.")
+
+	_, err := a.Run(context.Background(), "test")
+	if err == nil {
+		t.Fatal("expected error, got nil")
+	}
+}
+
+func TestAgent_EmptySystem(t *testing.T) {
+	model := newSimpleMockModel("no system prompt")
+	a := New(model, "") // Empty system prompt
+
+	result, err := a.Run(context.Background(), "test")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "no system prompt" {
+		t.Errorf("unexpected result: %q", result)
+	}
+}
--- a/v2/agent/example_test.go
+++ b/v2/agent/example_test.go
@@ -0,0 +1,107 @@
+package agent_test
+
+import (
+	"context"
+	"fmt"
+	"os"
+
+	llm "gitea.stevedudenhoeffer.com/steve/go-llm/v2"
+	"gitea.stevedudenhoeffer.com/steve/go-llm/v2/agent"
+	"gitea.stevedudenhoeffer.com/steve/go-llm/v2/tools"
+)
+
+// A researcher agent that can search the web and browse pages.
+func Example_researcher() {
+	model := llm.OpenAI(os.Getenv("OPENAI_API_KEY")).Model("gpt-4o")
+
+	researcher := agent.New(model,
+		"You are a research assistant. Use web search to find information, "+
+			"then use the browser to read full articles when needed. "+
+			"Provide a concise summary of your findings.",
+		agent.WithTools(llm.NewToolBox(
+			tools.WebSearch(os.Getenv("BRAVE_API_KEY")),
+			tools.Browser(),
+		)),
+		agent.WithRequestOptions(llm.WithTemperature(0.3)),
+	)
+
+	result, err := researcher.Run(context.Background(), "What are the latest developments in Go generics?")
+	if err != nil {
+		fmt.Println("Error:", err)
+		return
+	}
+	fmt.Println(result)
+}
+
+// A coder agent that can read, write, and execute code.
+func Example_coder() {
+	model := llm.OpenAI(os.Getenv("OPENAI_API_KEY")).Model("gpt-4o")
+
+	coder := agent.New(model,
+		"You are a coding assistant. You can read files, write files, and execute commands. "+
+			"When asked to create a program, write the code to a file and then run it to verify it works.",
+		agent.WithTools(llm.NewToolBox(
+			tools.ReadFile(),
+			tools.WriteFile(),
+			tools.Exec(
+				tools.WithAllowedCommands([]string{"go", "python", "node", "cat", "ls"}),
+				tools.WithWorkDir(os.TempDir()),
+			),
+		)),
+	)
+
+	result, err := coder.Run(context.Background(),
+		"Create a Go program that prints the first 10 Fibonacci numbers. Save it and run it.")
+	if err != nil {
+		fmt.Println("Error:", err)
+		return
+	}
+	fmt.Println(result)
+}
+
+// An orchestrator agent that delegates to specialized sub-agents.
+// The orchestrator breaks a complex task into subtasks and dispatches them
+// to the appropriate sub-agent via tool calls.
+func Example_orchestrator() {
+	model := llm.OpenAI(os.Getenv("OPENAI_API_KEY")).Model("gpt-4o")
+
+	// Specialized sub-agents
+	researcher := agent.New(model,
+		"You are a research assistant. Search the web for information on the given topic "+
+			"and return a concise summary.",
+		agent.WithTools(llm.NewToolBox(
+			tools.WebSearch(os.Getenv("BRAVE_API_KEY")),
+		)),
+	)
+
+	coder := agent.New(model,
+		"You are a coding assistant. Write and test code as requested. "+
+			"Save files and run them to verify they work.",
+		agent.WithTools(llm.NewToolBox(
+			tools.ReadFile(),
+			tools.WriteFile(),
+			tools.Exec(tools.WithAllowedCommands([]string{"go", "python"})),
+		)),
+	)
+
+	// Orchestrator can delegate to both sub-agents
+	orchestrator := agent.New(model,
+		"You are a project manager. Break complex tasks into research and coding subtasks. "+
+			"Use delegate_research for information gathering and delegate_coding for implementation. "+
+			"Synthesize the results into a final answer.",
+		agent.WithTools(llm.NewToolBox(
+			researcher.AsTool("delegate_research",
+				"Delegate a research task. Provide a clear question or topic to research."),
+			coder.AsTool("delegate_coding",
+				"Delegate a coding task. Provide clear requirements for what to implement."),
+		)),
+	)
+
+	result, err := orchestrator.Run(context.Background(),
+		"Research how to implement a binary search tree in Go, then create one with insert and search operations.")
+	if err != nil {
+		fmt.Println("Error:", err)
+		return
+	}
+	fmt.Println(result)
+}
--- a/v2/constructors.go
+++ b/v2/constructors.go
@@ -16,7 +16,7 @@ func OpenAI(apiKey string, opts ...ClientOption) *Client {
 	for _, opt := range opts {
 		opt(cfg)
 	}
-	return newClient(openaiProvider.New(apiKey, cfg.baseURL))
+	return NewClient(openaiProvider.New(apiKey, cfg.baseURL))
 }

 // Anthropic creates an Anthropic client.
@@ -30,7 +30,7 @@ func Anthropic(apiKey string, opts ...ClientOption) *Client {
 		opt(cfg)
 	}
 	_ = cfg // Anthropic doesn't support custom base URL in the SDK
-	return newClient(anthProvider.New(apiKey))
+	return NewClient(anthProvider.New(apiKey))
 }

 // Google creates a Google (Gemini) client.
@@ -44,5 +44,5 @@ func Google(apiKey string, opts ...ClientOption) *Client {
 		opt(cfg)
 	}
 	_ = cfg // Google doesn't support custom base URL in the SDK
-	return newClient(googleProvider.New(apiKey))
+	return NewClient(googleProvider.New(apiKey))
 }
--- a/v2/llm.go
+++ b/v2/llm.go
@@ -12,8 +12,9 @@ type Client struct {
 	middleware []Middleware
 }

-// newClient creates a Client backed by the given provider.
-func newClient(p provider.Provider) *Client {
+// NewClient creates a Client backed by the given provider.
+// Use this to integrate custom provider implementations or for testing.
+func NewClient(p provider.Provider) *Client {
 	return &Client{p: p}
 }

--- a/v2/model_test.go
+++ b/v2/model_test.go
@@ -117,7 +117,7 @@ func TestModel_Complete_WithTools(t *testing.T) {

 func TestClient_Model(t *testing.T) {
 	mp := newMockProvider(provider.Response{Text: "hi"})
-	client := newClient(mp)
+	client := NewClient(mp)
 	model := client.Model("test-model")

 	resp, err := model.Complete(context.Background(), []Message{UserMessage("test")})
@@ -144,7 +144,7 @@ func TestClient_WithMiddleware(t *testing.T) {
 	}

 	mp := newMockProvider(provider.Response{Text: "ok"})
-	client := newClient(mp).WithMiddleware(mw)
+	client := NewClient(mp).WithMiddleware(mw)
 	model := client.Model("test-model")

 	_, err := model.Complete(context.Background(), []Message{UserMessage("test")})