Building Production Agentic AI: MCP, Tools, and Real-World Integration

Most AI demos stop at chat. But real agentic systems need to actually do things—call APIs, read databases, trigger workflows. This talk is about building agents that work in production, not just in notebooks.

I'll walk through how to build reliable agentic systems, focusing on three things that matter: getting agents to use tools correctly, handling the inevitable failures when APIs are down or rate-limited, and scaling beyond toy examples to real workloads.

MCP is a big part of this story. Instead of hardcoding every tool your agent might need, MCP lets you expose capabilities dynamically. I'll show how to build MCP servers that plug into agent workflows, how they compare to function calling, and when each approach makes sense.

Expect lots of code, real trade-offs, and things I learned the hard way deploying this stuff.

You'll understand the building blocks—reasoning loops, tool orchestration, memory—and when to use function calling vs MCP vs building custom integrations.

Learn how to expose tools through MCP, handle auth and rate limits properly, and make your servers robust enough for production use.

When to use RAG, when to summarize, and how to balance what your agent remembers with token costs and performance.