Coding Agents

Coding agents use LLMs to generate, review, test, and refine code through multi-step workflows. Unlike simple code generation prompts, coding agents maintain context across iterations, execute code, and correct errors automatically.

Core Pattern: Generate → Execute → Refine

The fundamental coding agent loop:

1. Generate — The model writes code based on a specification
2. Execute — Run the code in a sandboxed environment
3. Observe — Capture output, errors, and test results
4. Refine — Feed results back to the model for corrections
5. Repeat until tests pass or the step limit is reached

package main

import (
    "context"
    "fmt"
    "log"
    "os"
    "os/exec"
    "strings"

    "github.com/digitallysavvy/go-ai/pkg/agent"
    "github.com/digitallysavvy/go-ai/pkg/provider/types"
    "github.com/digitallysavvy/go-ai/pkg/providers/anthropic"
)

func main() {
    ctx := context.Background()

    p := anthropic.New(anthropic.Config{APIKey: os.Getenv("ANTHROPIC_API_KEY")})
    model, _ := p.LanguageModel(anthropic.ClaudeSonnet4_5)

    // Build a coding agent with shell execution tools
    codingAgent := agent.NewToolLoopAgent(agent.AgentConfig{
        Model: model,
        System: `You are an expert Go programmer. Your job is to write correct, idiomatic Go code.

Workflow:
1. Write the requested code
2. Use the write_file tool to save it
3. Use the run_go tool to test it
4. Fix any errors and repeat until the code passes
5. When the code passes, report the final solution`,

        Tools: []types.Tool{
            writeFileTool(),
            runGoTool(),
        },
        MaxSteps: 15,
    })

    result, err := codingAgent.Execute(ctx,
        "Write a Go function `ReverseWords(s string) string` that reverses the order of words in a sentence. Include a test that verifies it.")
    if err != nil {
        log.Fatalf("Agent failed: %v", err)
    }

    fmt.Println("=== Final Result ===")
    fmt.Println(result.Text)
    fmt.Printf("\nCompleted in %d steps\n", len(result.Steps))
}

// writeFileTool writes content to a file in a temp directory.
func writeFileTool() types.Tool {
    return types.Tool{
        Name:        "write_file",
        Description: "Write content to a file. Use this to save generated Go code.",
        Parameters: map[string]interface{}{
            "type": "object",
            "properties": map[string]interface{}{
                "path":    map[string]interface{}{"type": "string", "description": "File path (relative to /tmp/coding-agent/)"},
                "content": map[string]interface{}{"type": "string", "description": "File content"},
            },
            "required": []string{"path", "content"},
        },
        Execute: func(ctx context.Context, input map[string]interface{}, opts types.ToolExecutionOptions) (interface{}, error) {
            path := "/tmp/coding-agent/" + input["path"].(string)
            content := input["content"].(string)
            os.MkdirAll("/tmp/coding-agent", 0755)
            if err := os.WriteFile(path, []byte(content), 0644); err != nil {
                return nil, fmt.Errorf("failed to write file: %w", err)
            }
            return map[string]interface{}{"status": "ok", "path": path}, nil
        },
    }
}

// runGoTool runs `go test` or `go run` in the temp directory.
func runGoTool() types.Tool {
    return types.Tool{
        Name:        "run_go",
        Description: "Run `go test ./...` or `go run <file>` in the working directory. Returns stdout, stderr, and exit code.",
        Parameters: map[string]interface{}{
            "type": "object",
            "properties": map[string]interface{}{
                "command": map[string]interface{}{
                    "type":        "string",
                    "description": "The go command to run: 'test', 'run <file>', or 'build'",
                },
            },
            "required": []string{"command"},
        },
        Execute: func(ctx context.Context, input map[string]interface{}, opts types.ToolExecutionOptions) (interface{}, error) {
            command := input["command"].(string)
            parts := strings.Fields("go " + command)

            cmd := exec.CommandContext(ctx, parts[0], parts[1:]...)
            cmd.Dir = "/tmp/coding-agent"

            out, err := cmd.CombinedOutput()
            exitCode := 0
            if err != nil {
                if exitErr, ok := err.(*exec.ExitError); ok {
                    exitCode = exitErr.ExitCode()
                }
            }

            return map[string]interface{}{
                "output":    string(out),
                "exit_code": exitCode,
                "success":   exitCode == 0,
            }, nil
        },
    }
}

Code Review Agent

A review agent examines existing code and suggests improvements without needing to execute anything:

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/digitallysavvy/go-ai/pkg/ai"
    "github.com/digitallysavvy/go-ai/pkg/provider/types"
    "github.com/digitallysavvy/go-ai/pkg/providers/openai"
)

type ReviewResult struct {
    Issues       []CodeIssue
    Suggestions  []string
    OverallScore int // 1-10
}

type CodeIssue struct {
    Severity    string // "critical", "warning", "info"
    Line        int
    Description string
    Fix         string
}

func reviewCode(ctx context.Context, code string) (*ReviewResult, error) {
    p := openai.New(openai.Config{APIKey: os.Getenv("OPENAI_API_KEY")})
    model, _ := p.LanguageModel("gpt-4o")

    result, err := ai.GenerateText(ctx, ai.GenerateTextOptions{
        Model: model,
        System: `You are an expert Go code reviewer. Analyze code for:
- Security vulnerabilities (SQL injection, command injection, etc.)
- Memory leaks and resource management issues
- Race conditions and concurrency bugs
- Performance problems
- Style and idiomatic Go violations
Return a structured JSON review.`,
        Messages: []types.Message{
            {
                Role: types.RoleUser,
                Content: []types.ContentPart{
                    types.TextContent{Text: fmt.Sprintf("Review this Go code:\n```go\n%s\n```", code)},
                },
            },
        },
        Output: ai.ObjectOutput[ReviewResult](ai.ObjectOutputOptions{
            Schema:      ai.SchemaFor[ReviewResult](),
            Name:        "code_review",
            Description: "Structured code review with issues and suggestions",
        }),
    })
    if err != nil {
        return nil, err
    }

    review, ok := result.Output.(ReviewResult)
    if !ok {
        return nil, fmt.Errorf("unexpected output type")
    }
    return &review, nil
}

func main() {
    ctx := context.Background()

    code := `
func processUserInput(db *sql.DB, username string) error {
    query := "SELECT * FROM users WHERE name = '" + username + "'"
    rows, err := db.Query(query)
    if err != nil {
        return err
    }
    defer rows.Close()
    return nil
}`

    review, err := reviewCode(ctx, code)
    if err != nil {
        log.Fatal(err)
    }

    fmt.Printf("Overall Score: %d/10\n\n", review.OverallScore)
    for _, issue := range review.Issues {
        fmt.Printf("[%s] Line %d: %s\n  Fix: %s\n\n",
            issue.Severity, issue.Line, issue.Description, issue.Fix)
    }
}

Multi-Step Code Refinement

Some tasks require multiple rounds of generation and feedback. Use GenerateText directly for fine-grained control:

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/digitallysavvy/go-ai/pkg/ai"
    "github.com/digitallysavvy/go-ai/pkg/provider"
    "github.com/digitallysavvy/go-ai/pkg/provider/types"
    "github.com/digitallysavvy/go-ai/pkg/providers/anthropic"
)

// CodeRefinementSession manages a multi-step code refinement workflow.
type CodeRefinementSession struct {
    model    provider.LanguageModel
    history  []types.Message
    maxRounds int
}

func NewCodeRefinementSession(model provider.LanguageModel, maxRounds int) *CodeRefinementSession {
    return &CodeRefinementSession{model: model, maxRounds: maxRounds}
}

// Generate produces or refines code based on the current conversation state.
func (s *CodeRefinementSession) Generate(ctx context.Context, input string) (string, error) {
    s.history = append(s.history, types.Message{
        Role:    types.RoleUser,
        Content: []types.ContentPart{types.TextContent{Text: input}},
    })

    result, err := ai.GenerateText(ctx, ai.GenerateTextOptions{
        Model:    s.model,
        System:   "You are an expert Go programmer. Write production-quality, well-documented Go code.",
        Messages: s.history,
    })
    if err != nil {
        return "", err
    }

    s.history = append(s.history, types.Message{
        Role:    types.RoleAssistant,
        Content: []types.ContentPart{types.TextContent{Text: result.Text}},
    })

    return result.Text, nil
}

func main() {
    ctx := context.Background()

    p := anthropic.New(anthropic.Config{APIKey: os.Getenv("ANTHROPIC_API_KEY")})
    model, _ := p.LanguageModel(anthropic.ClaudeSonnet4_5)

    session := NewCodeRefinementSession(model, 3)

    // Round 1: Initial implementation
    code, err := session.Generate(ctx, "Implement a concurrent worker pool in Go that processes jobs from a channel.")
    if err != nil {
        log.Fatal(err)
    }
    fmt.Println("Round 1 - Initial implementation:")
    fmt.Println(code)

    // Round 2: Add error handling
    code, err = session.Generate(ctx, "Add proper error handling so failed jobs don't silently fail. Workers should send errors to an error channel.")
    if err != nil {
        log.Fatal(err)
    }
    fmt.Println("\nRound 2 - With error handling:")
    fmt.Println(code)

    // Round 3: Add graceful shutdown
    code, err = session.Generate(ctx, "Add graceful shutdown so the pool waits for all in-progress jobs to complete before exiting.")
    if err != nil {
        log.Fatal(err)
    }
    fmt.Println("\nRound 3 - With graceful shutdown:")
    fmt.Println(code)
}

Streaming Code Generation

For interactive code generation (showing code as it's written), use StreamText:

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/digitallysavvy/go-ai/pkg/ai"
    "github.com/digitallysavvy/go-ai/pkg/providers/openai"
)

func main() {
    ctx := context.Background()

    p := openai.New(openai.Config{APIKey: os.Getenv("OPENAI_API_KEY")})
    model, _ := p.LanguageModel("gpt-4o")

    result, err := ai.StreamText(ctx, ai.StreamTextOptions{
        Model: model,
        System: "You are an expert Go programmer. Write clean, idiomatic code with inline comments.",
        Prompt: "Implement a simple HTTP rate limiter middleware for a Go web server using the token bucket algorithm.",
    })
    if err != nil {
        log.Fatal(err)
    }
    defer result.Close()

    fmt.Println("Generating rate limiter implementation...")
    fmt.Println()

    // Print code tokens as they arrive
    for chunk := range result.Chunks() {
        fmt.Print(chunk.Text)
    }
    fmt.Println()
}

Best Practices

1. Sandbox Execution

Always run generated code in an isolated environment:

// Use Docker for strong isolation
cmd := exec.CommandContext(ctx,
    "docker", "run", "--rm",
    "--memory=256m",       // Memory limit
    "--cpus=0.5",          // CPU limit
    "--network=none",      // No network access
    "--read-only",         // Read-only filesystem
    "-v", "/tmp/code:/workspace:rw",
    "golang:1.22",
    "sh", "-c", "cd /workspace && go test ./...",
)

2. Timeout Control

Always set timeouts on execution:

ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
defer cancel()

result, err := codingAgent.Execute(ctx, specification)
if err != nil {
    if ctx.Err() == context.DeadlineExceeded {
        log.Println("Code generation timed out")
    }
}

3. Clear Specifications

The quality of generated code directly reflects the quality of the specification:

// Good — specific, testable requirements
specification := `
Implement a Go function with this signature:
    func ParseDuration(s string) (time.Duration, error)

Requirements:
- Accept formats: "1h30m", "90m", "1.5h", "5400s"
- Return an error for invalid formats
- Include table-driven tests covering valid inputs, edge cases, and invalid inputs
`

// Bad — vague
specification := "Write a duration parser"

4. Iterative Feedback

Include specific error context when feeding back failures:

feedbackPrompt := fmt.Sprintf(`The code failed with:

Test output:
%s

Please fix the failing tests. Focus on:
1. The exact error messages above
2. Don't change the function signature
3. Don't modify the test file`, testOutput)

5. Context Window Management

For long coding sessions, manage token usage:

// Use summarization for long refinement sessions
if len(session.history) > 20 {
    summary, _ := session.Generate(ctx,
        "Summarize the key requirements and design decisions from our discussion so far in 2-3 paragraphs.")
    // Reset history with just the summary
    session.history = []types.Message{
        {Role: types.RoleSystem, Content: []types.ContentPart{
            types.TextContent{Text: "Previous context: " + summary},
        }},
    }
}

See Also

• Building Agents — agent configuration and patterns
• Loop Control — StopWhen and MaxSteps
• Memory Management — managing conversation history
• Tools and Tool Calling — building execution tools
• Generating Structured Data — typed code review results