Lecture Notes: Practical Guide on Building AI Agents

Introduction

• OpenAI released a practical guide on building agents.
• 2025 is anticipated to be 'the year of agents'.
• Many major labs, including Anthropic and Google, have their own guides on agents.
• A talk was given at Google Next on building AI agents.

Current Landscape

• Various SDKs and libraries available for building agents: OpenAI, Google, Langraph, Crew AI.
• Important to keep agentic systems simple; think of agents as LLMs with function call capabilities.

Definition of an Agent

• An agent is an AI-driven application with:
  • A model or LLM at its core.
  • Access to tools for enhanced capabilities.
  • An orchestration layer for operation.
• Agents autonomously accomplish tasks.

OpenAI's Practical Guide

Components of Agentic Systems

1. Model (LLM): Powers reasoning and decision-making.
2. Tools: Expands capabilities through function calls (data, action, orchestration).
3. Instructions: System instructions that guide behavior.

Choosing Models

• Not every application needs the most capable model.
• Optimize for performance, cost, and latency.
• Use the most capable model for initial prototypes, then optimize.

Tools Classification

• Data Tools: For retrieving context and information.
• Action Tools: To interact with systems (e.g., updating databases).
• Orchestration Tools: Agents serving as tools to other agents.

Instructions

• Provide clear, detailed guidelines to the LLM.
• Use existing documentation for context.
• Break down tasks and capture edge cases.

Orchestration Layer

• Single-Agent Systems: Executes workflows in a loop.
• Multi-Agent Systems: Divides tasks among multiple agents.
• Patterns:
  • Manager pattern (central orchestrator agent).
  • Decentralized pattern (autonomous agents with control handoffs).

Guardrails

• Important for managing data privacy and safety.
• Should operate independently of the agent system.

Implementation Tips

• Use simple Python functions for defining tools and guardrails.
• Limit the number of tools per agent (preferably under 10).
• Iterative refinement is crucial: update based on observed interactions and failures.

Practical Considerations

• Validate the need for an agentic solution based on complexity, rules, and unstructured data.
• Start with single-agent systems; consider multi-agent frameworks for complex applications.

Key Recommendations

• Focus on iterative refinements and evolve datasets.
• Measure application-specific metrics, such as accuracy or recall.
• Keep adapting guidelines based on real-world use cases and challenges.

Conclusion

• Convergence of ideas among labs like OpenAI, Google, and Anthropic towards industry standards.
• Excitement about the future development of agentic systems.