New York - Agentic AI

Ever wondered how AI coding agents like Claude Code actually work under the hood?

In this live build, we’ll create a practical coding agent from scratch with no agent framework, and no magic. We’ll build the tool-use loop that lets an LLM work with a local project while keeping you in control.

The agent will be able to read files, write and edit code, run shell commands, and ask permission.

This is not an attempt to recreate every feature of Claude Code. It is a hands-on look at the underlying pattern, and a way to understand how these systems actually work by building one.

We’ll explore:

• Setting up the LLM client and the agent’s basic loop
• Defining tools to read files, write and edit code, and run shell commands
• Executing tool calls and keeping the agent’s state straight
• Building a terminal REPL loop that works through a natural-language task
• Adding permission prompts before the agent takes action
• Handling validation, errors, cleanup, and the final handoff

By the end, you’ll have a working terminal-based agent that can read a small codebase, find and fix a bounded issue, run a check, report what it did, and more.

This is a builder-first session for developers, automation enthusiasts, and anyone who wants to move from using coding agents to understanding how to design them.

What happens when you stop typing every instruction and start directing the work out loud?

This is a live ride with ChatGPT Voice in Codex. We will start with a direction, follow the work where it goes, interrupt when it needs a better turn, and see what useful things emerge.

No polished slide deck. No rigid script. Just real work, spoken out loud, with room for detours, discoveries, and course corrections.

Ask. Interrupt. Redirect. Research. Build. Live.

Come see what changes when the conversation becomes part of the workflow.