Office Hours.
Agenda Concepts Demo Build Paths Resources
Session 01 · April 2, 2026

We kicked off Office-Hours.dev at 1900 Broadway Tower—Oakland yesterday.

Remember the first time you tried an iPod? Or rode in a self-driving car? Building with AI agents has a similar magical quality.

Ranging from first-time users to multi-agent orchestrators, the group brought the energy and plenty of magic. So many stayed well after our 3-hour session ended, hyped to share their passions, goals, and vision for the future.

I'm very fortunate to have a great team and community to lean on. Humbled to be across such amazing projects, and have the opportunity to contribute some of my learnings from along the way. Feeling inspired heading into the weekend.

Session 02: Agent Coordination and Orchestration — April 22, 2026

🙏 Adeniji Asabi, Brian💥 Sparkes, Colin Behring, Jason Esguerra, Ryan George, and Ted Yamada-Dessert for helping bring the series to life.

— Andrew Smith

Session 01

Agent First.

AI gets useful when it can interact with the outside world. Today we connect models to real tools, real APIs, and real data. You'll leave with a working agent.

office-hours.dev · 1900 Broadway, 2nd Floor · Oakland, CA

Agenda

How the next 3 hours work

12:00

Welcome + pulse check

Who's in the room. What you're hoping to build. Quick show of hands on experience level.

12:10

What makes an agent an agent?

The shift from chatbot to agent. Why tool use changes everything. The 3 things that matter.

12:25

Core concepts: tools, function calling, MCP

The building blocks. What they are, how they work, when to use which one.

12:40

Live demo: build an agent together

From zero to working agent in 15 minutes. Everyone follows along or watches.

12:55

Choose your path

Pick a build project based on your skill level and what you care about. We help you scope it.

1:00

Build session (2 hours)

You build. We float. Ask anything. Change direction. Go deeper. The room is yours.

2:45

Show + tell (optional)

Quick demos of what people built. 2 minutes each. No pressure, but encouraged.

Pulse Check

Who's in the room?

Quick show of hands. No wrong answers.

01

Have you used ChatGPT, Claude, or Gemini?

Just the chat interface. Asking questions, writing, brainstorming.

02

Have you written code with AI help?

Used Copilot, Cursor, Claude Code, Codex, or similar.

03

Have you built an agent that uses tools?

Connected a model to an API, database, or external service.

04

Have you shipped an agent to production?

Running in the real world, handling real data or real money.

The Shift

Chatbot vs. agent

Chatbot

You ask a question. It answers with text.

All it can do is generate words.

No memory between sessions.

Can't take action in the world.

Knowledge frozen at training cutoff.

Agent

You give it a goal. It figures out the steps.

It can call tools, APIs, databases.

It can maintain context and state.

It can take real actions: send emails, query data, make purchases.

It can access live, current information.

The key insight: An agent is just a language model with the ability to use tools. That's it. The model decides when to use a tool, which tool to use, and what to do with the result. Everything else is engineering.

The Building Blocks

Three things that make an agent work

01

Tool definitions

You describe what tools are available. Name, description, parameters. The model reads these to decide what it can do.

02

Function calling

The model outputs a structured request: "call this function with these arguments." Your code executes it and returns the result.

03

The loop

Model thinks → calls tool → gets result → thinks again. This loop continues until the task is done. That's the agent.

MCP

Model Context Protocol

MCP is a standard way to connect AI models to tools and data sources. Think of it as USB for AI. One protocol, any tool, any model.

Without MCP

Every tool needs a custom integration for every model. N tools × M models = N×M integrations.

With MCP

Every tool implements one standard. Every model speaks one protocol. N + M integrations total.

Why this matters today: MCP servers already exist for Google Drive, Slack, GitHub, databases, Stripe, and hundreds more. You don't have to build integrations from scratch. You connect to what already exists.

How It Works

Anatomy of a tool call

Here's what actually happens when an agent uses a tool:

// 1. You define the tool { "name": "get_weather", "description": "Get current weather for a city", "parameters": { "city": { "type": "string" } } } // 2. The model decides to use it → Model output: tool_use("get_weather", { city: "Oakland" }) // 3. Your code executes the function → API call to weather service → returns { temp: 68, condition: "sunny" } // 4. Result goes back to the model → Model: "It's 68°F and sunny in Oakland right now."

The model never calls the API directly. It outputs a structured request. Your code is the one making the actual call. You control what happens.

Live Demo

Let's build one together

We're going to build a simple agent from scratch. Follow along on your laptop, or just watch. We'll go step by step.

What we're building

An agent that can search the web, summarize what it finds, and answer follow-up questions using the results.

What you need open

A browser with Claude.ai, ChatGPT, or your preferred model. That's it for the demo. API keys come later.

We'll use Claude for the demo, but everything we cover works across models. The concepts are the same whether you're using Claude, GPT, Gemini, or open-source models.

Approaches

Three ways to build agents today

A

Chat interface + MCP

Use Claude Desktop, Cursor, or ChatGPT with MCP servers connected. No code required. The model calls tools through the interface.

Best for: getting started fast

B

Claude Code / Codex / Cursor

AI-powered coding tools that are themselves agents. They read your codebase, write code, run commands, and iterate. You direct, they build.

Best for: building software with AI

C

API + custom code

Call the model API directly. Define tools in code. Build your own agent loop. Full control over every step.

Best for: production systems

Build Session

Choose your path

Pick one. Build it in the next 2 hours. We're here to help.

New to AI

Personal research agent

Build an agent in Claude or ChatGPT that can search the web, read documents, and answer questions about topics you care about.

  • Connect MCP servers (web search, file reading)
  • Give it a system prompt with your interests
  • Ask it to research something real
  • Refine until the output is actually useful

Tools: Claude.ai or ChatGPT · No code required

New to AI

Daily briefing agent

Build an agent that pulls information from multiple sources and gives you a personalized morning briefing.

  • Connect to news, weather, calendar
  • Define what you want to know each morning
  • Get a single, clean summary
  • Customize the format and tone

Tools: Claude.ai or ChatGPT · No code required

Some experience

Slack / email assistant

Build an agent that monitors a channel or inbox and takes action: summarizes threads, drafts responses, flags urgent items.

  • Connect Slack or Gmail MCP server
  • Define rules for what matters
  • Auto-draft responses for review
  • Summarize long threads on demand

Tools: Claude Desktop + MCP · Light config

Some experience

Code review agent

Build an agent that reads a GitHub repo, understands the codebase, and gives meaningful feedback on pull requests.

  • Connect GitHub MCP server
  • Point it at a real repo
  • Ask it to review recent changes
  • Refine its review criteria

Tools: Claude Code or Cursor · Some coding

Ready to build

Custom tool agent via API

Build a custom agent loop using the Anthropic or OpenAI API. Define your own tools. Handle the full cycle: prompt → tool call → result → response.

  • Set up API keys via Keyfree
  • Define 2-3 custom tools
  • Build the agent loop in Python or Node
  • Handle errors, retries, multi-step reasoning

Tools: API + your editor · Full code

Ready to build

Build your own MCP server

Create an MCP server that wraps an API you use. Make it available to any MCP-compatible client. Ship a tool others can use.

  • Pick an API you know well
  • Implement the MCP server protocol
  • Define tools, resources, and prompts
  • Test it with Claude Desktop

Tools: Python or Node · Full code

API Keys

Need API access? Use Keyfree.

If your project needs API keys (Anthropic, OpenAI, Google, etc.), we'll provision them on-site through Keyfree. No credential management required. You use the capability without seeing the key.

keyfree.dev

Resources

References for the build session

Claude

docs.anthropic.com/en/docs/build-with-claude/tool-use

modelcontextprotocol.io

docs.anthropic.com/en/docs/claude-code

OpenAI

platform.openai.com/docs/guides/function-calling

platform.openai.com/docs/guides/tools

MCP Servers

github.com/modelcontextprotocol/servers

MCP server directory: mcp.so

This session

office-hours.dev

Recording will be posted after the session

2:45 PM

Show + tell

Built something? Show the room. 2 minutes each. No slides needed. Just share your screen and tell us what you made.

Not required. But you'll be surprised how much you built in 2 hours.

What's Next

Session 02: Agent Coordination + Orchestration

One agent is useful. Multiple agents working together is a system. April 22, same time, same place.

Also this month

April 9 · The Infrastructure Strikes Back

Red Team vs. Blue Team adversarial experiment

April 16 · ClawCamp

Free for first 100 RSVPs

Stay connected

office-hours.dev

hlos.ai

day-zero.dev

Office Hours · Session 01 · Agent First · April 2, 2026 · Oakland, CA · office-hours.dev