react

ReAct Agent Framework

A production-ready implementation of ReAct (Reasoning and Acting) agents with advanced cognitive architectures including self-reflection, task planning, and memory optimization with forgetting curves.

Architecture Overview

┌─────────────────────────────────────────────────────────────────┐
│                         ReAct Agent                             │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌───────────────┐  ┌──────────────┐  ┌──────────────────────┐  │
│  │ Execution     │  │ Task         │  │ Self                 │  │
│  │ Modes         │  │ Planning     │  │ Reflection           │  │
│  ├───────────────┤  ├──────────────┤  ├──────────────────────┤  │
│  │ • ReAct       │  │ • Decompose  │  │ • Strategy Analysis  │  │
│  │ • ReWOO       │  │ • Templates  │  │ • Performance        │  │
│  │ • Hybrid      │  │ • Parallel   │  │ • Learning           │  │
│  └───────────────┘  └──────────────┘  └──────────────────────┘  │
│                                                                 │
│  ┌───────────────┐  ┌──────────────┐  ┌──────────────────────┐  │
│  │ Memory        │  │ Tool         │  │ Execution            │  │
│  │ Optimizer     │  │ Registry     │  │ History              │  │
│  ├───────────────┤  ├──────────────┤  ├──────────────────────┤  │
│  │ • Forgetting  │  │ • Dynamic    │  │ • Tracking           │  │
│  │   Curve       │  │   Loading    │  │ • Metrics            │  │
│  │ • Compression │  │ • Validation │  │ • Analysis           │  │
│  │ • Indexing    │  │ • Execution  │  │ • Interceptors       │  │
│  └───────────────┘  └──────────────┘  └──────────────────────┘  │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Core Features

🎯 Execution Modes

ReAct Mode (Classic)

Traditional ReAct loop with iterative Thought → Action → Observation cycles.

agent := react.NewReActAgent("my-agent", "Assistant",
    react.WithExecutionMode(react.ModeReAct))

• Best for: Exploratory tasks, interactive queries, complex reasoning
• Token usage: Standard (full reasoning traces)
• Flexibility: Maximum adaptability

ReWOO Mode (Plan-then-Execute)

Reasoning Without Observation - creates plan upfront then executes.

agent := react.NewReActAgent("my-agent", "Assistant",
    react.WithExecutionMode(react.ModeReWOO))

• Best for: Structured workflows, batch processing, predictable tasks
• Token usage: ~65% reduction vs ReAct
• Speed: Faster execution with less LLM calls

Hybrid Mode (Adaptive)

Automatically selects mode based on task complexity analysis.

agent := react.NewReActAgent("my-agent", "Assistant",
    react.WithExecutionMode(react.ModeHybrid))

• Best for: General-purpose applications, varying workloads
• Intelligence: Analyzes task keywords, length, and available tools
• Balance: Optimal token/performance trade-off

🧠 Memory Optimization

Implements Ebbinghaus forgetting curve with intelligent compression strategies:

optimizer := react.NewMemoryOptimizer(
    24*time.Hour,  // retention period
    0.3,           // forget threshold
)

Features:

• Forgetting Curve: R = e^(-t/S) where R is retention, t is time, S is strength
• Compression Strategies:
  • Summarization: Groups similar memories into summaries
  • Merging: Combines semantically similar items
  • Pruning: Removes low-importance memories
• Semantic Search: Embedding-based similarity matching
• Category Organization: Automatic task categorization
• Importance Scoring: Based on success, complexity, and access patterns

📊 Self-Reflection System

Multi-level reflection with pattern recognition and learning:

reflector := react.NewSelfReflector(
    3,                     // reflection depth
    100*time.Millisecond,  // reflection delay
)

Reflection Types:

• Strategy Reflection: Identifies repeated patterns and tool usage
• Performance Reflection: Analyzes success rates and efficiency
• Learning Reflection: Extracts reusable insights
• Error Reflection: Learns from failures

📋 Task Planning

Sophisticated planning with decomposition and parallelization:

planner := react.NewTaskPlanner(
    react.DecompositionFirst,  // or Interleaved
    5,                        // max decomposition depth
)

Features:

• Plan Templates: Reusable patterns for common tasks
• Task Decomposition: Breaks complex tasks into subtasks
• Dependency Analysis: Identifies parallel execution opportunities
• Critical Path: Marks essential vs optional steps
• Optimization: Reorders steps for efficiency

Quick Start

package main

import (
    "context"
    "github.com/XiaoConstantine/dspy-go/pkg/agents/react"
    "github.com/XiaoConstantine/dspy-go/pkg/core"
)

func main() {
    // Create agent with all advanced features
    agent := react.NewReActAgent("research-agent", "Research Assistant",
        react.WithExecutionMode(react.ModeHybrid),
        react.WithReflection(true, 3),
        react.WithPlanning(react.Interleaved, 5),
        react.WithMemoryOptimization(24*time.Hour, 0.3),
        react.WithMaxIterations(10),
        react.WithTimeout(5*time.Minute),
    )

    // Initialize with typed signature
    type Input struct {
        Task string `json:"task"`
    }
    type Output struct {
        Result string `json:"result"`
    }

    signature := core.NewTypedSignature[Input, Output]()
    err := agent.Initialize(llm, signature.ToLegacySignature())

    // Register tools
    tools := []core.Tool{searchTool, calculatorTool, summarizerTool}
    for _, tool := range tools {
        agent.RegisterTool(tool)
    }

    // Execute with interceptors for monitoring
    result, err := agent.ExecuteWithInterceptors(
        context.Background(),
        map[string]interface{}{"task": "Analyze market trends"},
        []core.AgentInterceptor{loggingInterceptor, metricsInterceptor},
    )

    // Analyze performance
    history := agent.GetExecutionHistory()
    reflections := agent.Reflector.GetTopReflections(3)
    metrics := agent.Reflector.GetMetrics()
}

Configuration

Complete Configuration Options

config := react.ReActAgentConfig{
    // Core Execution
    MaxIterations:    10,              // Max reasoning cycles
    ExecutionMode:    ModeHybrid,      // ReAct/ReWOO/Hybrid
    Timeout:         5*time.Minute,    // Overall timeout

    // Memory Management
    MemoryRetention:      24*time.Hour,  // How long to retain memories
    ForgetThreshold:      0.3,           // Minimum retention score
    EnableMemoryOpt:      true,          // Enable optimization
    CompressionThreshold: 100,           // Items before compression

    // Self-Reflection
    EnableReflection:     true,               // Enable reflection system
    ReflectionDepth:      3,                  // Analysis depth
    ReflectionDelay:      100*time.Millisecond, // Delay between reflections
    ReflectionThreshold:  0.7,                // Confidence threshold

    // Task Planning
    PlanningStrategy:     Interleaved,   // DecompositionFirst/Interleaved
    MaxPlanDepth:        5,              // Max decomposition depth
    EnablePlanning:      true,           // Enable planning system

    // Tool Execution
    ToolTimeout:         30*time.Second, // Per-tool timeout
    ParallelTools:       true,           // Enable parallel execution
    MaxToolRetries:      3,              // Retry failed tools

    // Monitoring
    EnableInterceptors:  true,           // Enable interceptor support
}

Advanced Usage

Memory Statistics and Analysis

// Get memory statistics
stats := agent.Optimizer.GetStatistics()
// Returns: total_items, categories, retention_rate, compression_ratio,
//          category_distribution, avg_importance, avg_access_count

// Find relevant memories
memories := agent.Optimizer.Retrieve(ctx, map[string]interface{}{
    "task": "previous research query",
})

Plan Metrics and Optimization

// Get plan metrics
plan, _ := agent.Planner.CreatePlan(ctx, input, tools)
metrics := agent.Planner.GetPlanMetrics(plan)
// Returns: total_steps, parallel_steps, critical_steps,
//          estimated_time, strategy, parallelization, tool_usage

// Validate plan
err := agent.Planner.ValidatePlan(plan)

Reflection Insights

// Get detailed metrics
metrics := agent.Reflector.GetMetrics()
// Returns: total_reflections, avg_confidence, success_rate,
//          avg_iterations, improvements, reflection_types

// Calculate improvement rate
improvement := agent.Reflector.CalculateImprovement()

// Reset for new session
agent.Reflector.Reset()

Performance Characteristics

Token Efficiency (GPT-3.5 Benchmarks)

Memory Performance

• Compression Ratio: 50-70% reduction in memory footprint
• Retrieval Speed: O(n log n) with category indexing
• Accuracy: 85%+ relevance for top-5 retrievals

Test Coverage

• Overall Coverage: 80.7%
• Core Components: 85%+
• Edge Cases: Comprehensive error handling

Tool Integration

Standard Tool Registration

// Implements core.Tool interface
type MyTool struct {
    name string
}

func (t *MyTool) Name() string { return t.name }
func (t *MyTool) Description() string { return "My custom tool" }
func (t *MyTool) Execute(ctx context.Context, params map[string]interface{}) (core.ToolResult, error) {
    // Implementation
}
// ... other required methods

agent.RegisterTool(&MyTool{})

Function Tools

tool := tools.NewFuncTool("search", "Search the web",
    schema,
    func(ctx context.Context, args map[string]interface{}) (*models.CallToolResult, error) {
        // Implementation
    })
agent.RegisterTool(tool)

Monitoring and Observability

Interceptor Support

// Built-in support for interceptors
agent.SetInterceptors([]core.AgentInterceptor{
    logging.NewInterceptor(),
    metrics.NewInterceptor(),
    tracing.NewInterceptor(),
})

// Or per-execution
result, err := agent.ExecuteWithInterceptors(ctx, input, interceptors)

Execution History

history := agent.GetExecutionHistory()
for _, record := range history {
    fmt.Printf("Task: %v\n", record.Input)
    fmt.Printf("Success: %v\n", record.Success)
    fmt.Printf("Actions taken: %d\n", len(record.Actions))
    fmt.Printf("Duration: %v\n", record.Duration)
}

Examples

See /examples/react_agent/ for complete examples including:

• Multi-tool orchestration
• All execution modes comparison
• Memory optimization demonstration
• Reflection analysis
• Performance monitoring

Testing

# Run all tests
go test ./pkg/agents/react/...

# Run with coverage
go test -cover ./pkg/agents/react/...

# Run benchmarks
go test -bench=. ./pkg/agents/react/

# Run specific test suites
go test -run TestMemoryOptimizer ./pkg/agents/react/
go test -run TestTaskPlanner ./pkg/agents/react/
go test -run TestSelfReflector ./pkg/agents/react/

Production Considerations

Resource Management

• Memory optimizer automatically manages memory footprint
• Configurable compression thresholds
• Automatic cleanup of old memories

Error Handling

• Comprehensive error recovery
• Atomic operations with rollback
• Critical vs non-critical step failures

Scalability

• Parallel tool execution support
• Efficient memory indexing
• Optimized dependency analysis

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE for details.