In a recent research publication: Building effective agents, Anthropic shared valuable insights about building effective Large Language Model (LLM) agents. What makes this research particularly interesting is its emphasis on simplicity and composability over complex frameworks. Let's explore how these principles translate into practical implementations using Spring AI.

While the pattern descriptions and diagrams are sourced from Anthropic's original publication, we'll focus on how to implement these patterns using Spring AI's features for model portability and structured output. We recommend reading the original paper first.

The agentic-patterns project implements the patterns discussed below.

Agentic Systems

The research publication makes an important architectural distinction between two types of agentic systems:

1. Workflows: Systems where LLMs and tools are orchestrated through predefined code paths (e.g., prescriptive system)
2. Agents: Systems where LLMs dynamically direct their own processes and tool usage

The key insight is that while fully autonomous agents might seem appealing, workflows often provide better predictability and consistency for well-defined tasks. This aligns perfectly with enterprise requirements where reliability and maintainability are crucial.

Let's examine how Spring AI implements these concepts through five fundamental patterns, each serving specific use cases:

1. Chain Workflow

The Chain Workflow pattern exemplifies the principle of breaking down complex tasks into simpler, more manageable steps.

When to Use:

• Tasks with clear sequential steps
• When you want to trade latency for higher accuracy
• When each step builds on the previous step's output

Here's a practical example from Spring AI's implementation:

public class ChainWorkflow {
    private final ChatClient chatClient;
    private final String[] systemPrompts;

    // Processes input through a series of prompts, where each step's output
    // becomes input for the next step in the chain.     
    public String chain(String userInput) {
        String response = userInput;
        for (String prompt : systemPrompts) {
            // Combine the system prompt with previous response
            String input = String.format("{%s}\n {%s}", prompt, response);
            // Process through the LLM and capture output
            response = chatClient.prompt(input).call().content();
        }
        return response;
    }
}

This implementation demonstrates several key principles:

• Each step has a focused responsibility
• Output from one step becomes input for the next
• The chain is easily extensible and maintainable

2. Parallelization Workflow

LLMs can work simultaneously on tasks and have their outputs aggregated programmatically. The parallelization workflow manifests in two key variations:

• Sectioning: Breaking tasks into independent subtasks for parallel processing
• Voting: Running multiple instances of the same task for consensus

When to Use:

• Processing large volumes of similar but independent items
• Tasks requiring multiple independent perspectives
• When processing time is critical and tasks are parallelizable

The Parallelization Workflow pattern demonstrates efficient concurrent processing of multiple Large Language Model (LLM) operations. This pattern is particularly useful for scenarios requiring parallel execution of LLM calls with automated output aggregation.

Here's a basic example of using the Parallelization Workflow:

List<String> parallelResponse = new ParallelizationWorkflow(chatClient)
    .parallel(
        "Analyze how market changes will impact this stakeholder group.",
        List.of(
            "Customers: ...",
            "Employees: ...",
            "Investors: ...",
            "Suppliers: ..."
        ),
        4
    );

This example demonstrates parallel processing of stakeholder analysis, where each stakeholder group is analyzed concurrently.

3. Routing Workflow

The Routing pattern implements intelligent task distribution, enabling specialized handling for different types of input.

This pattern is designed for complex tasks where different types of inputs are better handled by specialized processes. It uses an LLM to analyze input content and route it to the most appropriate specialized prompt or handler.

When to Use:

• Complex tasks with distinct categories of input
• When different inputs require specialized processing
• When classification can be handled accurately

Here's a basic example of using the Routing Workflow:

@Autowired
private ChatClient chatClient;

// Create the workflow
RoutingWorkflow workflow = new RoutingWorkflow(chatClient);

// Define specialized prompts for different types of input
Map<String, String> routes = Map.of(
    "billing", "You are a billing specialist. Help resolve billing issues...",
    "technical", "You are a technical support engineer. Help solve technical problems...",
    "general", "You are a customer service representative. Help with general inquiries..."
);

// Process input
String input = "My account was charged twice last week";
String response = workflow.route(input, routes);

4. Orchestrator-Workers

This pattern demonstrates how to implement more complex agent-like behavior while maintaining control:

• A central LLM orchestrates task decomposition
• Specialized workers handle specific subtasks
• Clear boundaries maintain system reliability

When to Use:

• Complex tasks where subtasks can't be predicted upfront
• Tasks requiring different approaches or perspectives
• Situations needing adaptive problem-solving

The implementation uses Spring AI's ChatClient for LLM interactions and consists of:

public class OrchestratorWorkersWorkflow {
    public WorkerResponse process(String taskDescription) {
        // 1. Orchestrator analyzes task and determines subtasks
        OrchestratorResponse orchestratorResponse = // ...

        // 2. Workers process subtasks in parallel
        List<String> workerResponses = // ...

        // 3. Results are combined into final response
        return new WorkerResponse(/*...*/);
    }
}

Usage Example:

ChatClient chatClient = // ... initialize chat client
OrchestratorWorkersWorkflow workflow = new OrchestratorWorkersWorkflow(chatClient);

// Process a task
WorkerResponse response = workflow.process(
    "Generate both technical and user-friendly documentation for a REST API endpoint"
);

// Access results
System.out.println("Analysis: " + response.analysis());
System.out.println("Worker Outputs: " + response.workerResponses());

5. Evaluator-Optimizer

The Evaluator-Optimizer pattern implements a dual-LLM process where one model generates responses while another provides evaluation and feedback in an iterative loop, similar to a human writer's refinement process. The pattern consists of two main components:

• Generator LLM: Produces initial responses and refines them based on feedback
• Evaluator LLM: Analyzes responses and provides detailed feedback for improvement

When to Use:

• Clear evaluation criteria exist
• Iterative refinement provides measurable value
• Tasks benefit from multiple rounds of critique

The implementation uses Spring AI's ChatClient for LLM interactions and consists of:

public class EvaluatorOptimizerWorkflow {
    public RefinedResponse loop(String task) {
        // 1. Generate initial solution
        Generation generation = generate(task, context);
        
        // 2. Evaluate the solution
        EvaluationResponse evaluation = evaluate(generation.response(), task);
        
        // 3. If PASS, return solution
        // 4. If NEEDS_IMPROVEMENT, incorporate feedback and generate new solution
        // 5. Repeat until satisfactory
        return new RefinedResponse(finalSolution, chainOfThought);
    }
}

Usage Example:

ChatClient chatClient = // ... initialize chat client
EvaluatorOptimizerWorkflow workflow = new EvaluatorOptimizerWorkflow(chatClient);

// Process a task
RefinedResponse response = workflow.loop(
    "Create a Java class implementing a thread-safe counter"
);

// Access results
System.out.println("Final Solution: " + response.solution());
System.out.println("Evolution: " + response.chainOfThought());

Spring AI's Implementation Advantages

Spring AI's implementation of these patterns offers several benefits that align with Anthropic's recommendations:

1. Model Portability

<!-- Easy model switching through dependencies -->
<dependency>
    <groupId>org.springframework.ai</groupId>
    <artifactId>spring-ai-openai-spring-boot-starter</artifactId>
</dependency>

2. Structured Output

// Type-safe handling of LLM responses
EvaluationResponse response = chatClient.prompt(prompt)
    .call()
    .entity(EvaluationResponse.class);

3. Consistent API

• Uniform interface across different LLM providers
• Built-in error handling and retries
• Flexible prompt management

Best Practices and Recommendations

Based on both Anthropic's research and Spring AI's implementations, here are key recommendations for building effective LLM-based systems:

• Start Simple

  • Begin with basic workflows before adding complexity
  • Use the simplest pattern that meets your requirements
  • Add sophistication only when needed

• Design for Reliability

  • Implement clear error handling
  • Use type-safe responses where possible
  • Build in validation at each step

• Consider Trade-offs

  • Balance latency vs. accuracy
  • Evaluate when to use parallel processing
  • Choose between fixed workflows and dynamic agents

Future Work

In Part 2 of this series, we'll explore how to build more advanced Agents that combine these foundational patterns with sophisticated features:

Pattern Composition

• Combining multiple patterns to create more powerful workflows
• Building hybrid systems that leverage the strengths of each pattern
• Creating flexible architectures that can adapt to changing requirements

Advanced Agent Memory Management

• Implementing persistent memory across conversations
• Managing context windows efficiently
• Developing strategies for long-term knowledge retention

Tools and Model-Context Protocol (MCP) Integration

• Leveraging external tools through standardized interfaces
• Implementing MCP for enhanced model interactions
• Building extensible agent architectures

Stay tuned for detailed implementations and best practices for these advanced features.

Tanzu Gen AI Solutions

VMware Tanzu Platform 10 Tanzu AI Server, powered by Spring AI provides:

• Enterprise-Grade AI Deployment: Production-ready solution for deploying AI applications within your VMware Tanzu environment
• Simplified Model Access: Streamlined access to Amazon Bedrock Nova models through a unified interface
• Security and Governance: Enterprise-level security controls and governance features
• Scalable Infrastructure: Built on Spring AI, the integration supports scalable deployment of AI applications while maintaining high performance

For more information about deploying AI applications with Tanzu AI Server, visit the VMware Tanzu AI documentation.

Conclusion

The combination of Anthropic's research insights and Spring AI's practical implementations provides a powerful framework for building effective LLM-based systems. By following these patterns and principles, developers can create robust, maintainable, and effective AI applications that deliver real value while avoiding unnecessary complexity.

The key is to remember that sometimes the simplest solution is the most effective. Start with basic patterns, understand your use case thoroughly, and only add complexity when it demonstrably improves your system's performance or capabilities.