<aside> 🦥

Sources

This is mainly from: https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf

</aside>

<aside> 🦥

Core Components

• Model: LLM powering reasoning and decision-making capabilities (GPT, Claude, Gemini, etc)
• Tools: External functions or APIs the agent uses to take actions
• Instructions: Guidelines and guardrails defining how the agent behaves </aside>

<aside> 🦥

Tools

• Data tools: APIs/MCPS, Query databases, read documents, search web
• Action tools: Add information to systems, update records, send messages
• Orchestration tools: Other agents that can be called to perform specific tasks </aside>

<aside> 🦥

🔌 How tools get connected: MCP

These days, tools are increasingly plugged in using MCP (Model Context Protocol): an open "USB-C for agents" standard. Build a tool once as an MCP server and any agent can use it. (See the What is MCP? note.)

</aside>

<aside> 🦥

Guardrails (I did not mention this in the video… whoops)

• Relevance classifier: Ensures responses stay within intended scope
• Safety classifier: Detects prompt injections and jailbreak attempts
• PII filter: Prevents exposure of personal identifiable information (PII)
• Tool safeguards: Monitors risk levels of available tools
• Rules-based protections: Blocklists, regex filters, input length limits
• Output validation: Ensures responses align with brand values and safety requirements </aside>

<aside> 🦥

Human Intervention

• Important for high-risk actions (payments, refunds, account changes)
• Needed when agent exceeds failure thresholds
• Critical during early deployment phases </aside>