​

Best Practices for Building Effective Agents

This guide outlines key principles and patterns for building successful AI agents, based on Anthropic’s research and experience working with teams across industries.

​

Core Principles

1. Start Simple
   • Begin with the simplest possible solution
   • Only increase complexity when demonstrably needed
   • Optimize single LLM calls with retrieval and in-context examples first
2. Choose the Right Architecture
   • Workflows: Use for predefined, well-structured tasks
   • Agents: Use when flexibility and model-driven decision-making are needed
   • Consider the tradeoffs between latency, cost, and task performance
3. Framework Considerations
   • Start by using LLM APIs directly when possible
   • Only add frameworks when they provide clear benefits
   • Ensure you understand the underlying code of any framework used

​

Common Patterns

​

1. Augmented LLM (Basic Building Block)

• Enhance LLMs with capabilities like:
  • Retrieval
  • Tools
  • Memory
• Focus on tailoring capabilities to specific use cases
• Ensure clear documentation for LLM interfaces

​

2. Workflow Patterns

​

Prompt Chaining

• Breaks tasks into sequential steps
• Best for tasks with clear, fixed subtasks
• Example: Generate content → Review → Translate

​

Routing

• Classifies inputs and directs to specialized handlers
• Useful for distinct categories requiring different approaches
• Example: Customer service query classification

​

Parallelization

• Sectioning: Breaking tasks into parallel subtasks
• Voting: Running multiple attempts for higher confidence
• Useful for speed optimization and quality assurance

​

Orchestrator-Workers

• Central LLM coordinates subtasks
• Dynamic task breakdown and delegation
• Best for complex, unpredictable workflows

​

Evaluator-Optimizer

• Iterative refinement through feedback loops
• Use when clear evaluation criteria exist
• Example: Complex content generation with quality checks

​

3. Autonomous Agents

• Use for open-ended problems
• Requires well-designed toolsets
• Include appropriate guardrails and stopping conditions

​

Tool Engineering Best Practices

1. Design Clear Interfaces
   • Write detailed documentation
   • Include example usage
   • Specify edge cases
   • Define clear boundaries between tools
2. Format Considerations
   • Allow sufficient tokens for model reasoning
   • Use formats familiar from internet text
   • Avoid complex formatting requirements
   • Minimize overhead like line counting or string escaping
3. Testing and Iteration
   • Test extensively with example inputs
   • Iterate based on model usage patterns
   • Design to prevent common mistakes
   • Use absolute paths over relative when possible

​

Further Reading

For more detailed information and examples, refer to the complete Anthropic research article on building effective agents.