A 5–6 hour hands-on workshop where you go from "what is an agent?" to shipping autonomous agents that read your inbox, review your PRs, organize your files, and run on a schedule. No theory-only sessions — every module ends with a working agent on your machine.
A prompt answers a question. A skill is a reusable expert behavior. An agent is a loop: it sets a goal, picks a tool, observes the result, and decides what to do next — until it's done. This workshop is about that loop, in practice.
Give the agent an outcome — not a script. It plans the steps. You set the budget and the boundaries.
Filesystem, web search, MCP servers, custom Python or TS functions. Each tool extends what the agent can actually do.
Act, look at the result, adjust, continue or stop. The whole craft is in shaping that loop and recovering from misfires.
This workshop assumes you've completed Claude Code Basics and Claude Code Skills, or you're already fluent at directing Claude in plain English and writing SKILL.md files. We move quickly into agent design from minute 30.
You don't need to write code from scratch — Claude does that. But because agents touch real systems (files, APIs, tools), you should be comfortable with: opening a terminal and running commands, editing a small YAML or JSON config file, and reading a short Python or JavaScript snippet to know roughly what it does. If you've ever cloned a repo, edited a config, or tweaked a script someone else wrote — you're ready. If those words feel foreign, take Basics first.
No slides past Module 1. You're at the keyboard from minute 30. Each module ends with a real artifact running on your machine.
The progression — prompt → skill → agent. The loop. Three ingredients: tools, memory, autonomy. When NOT to build an agent. Live demos comparing the three.
Build a custom subagent in .claude/agents/. Tools: Read, Write, Bash, Glob. Dry-run, then real run. Read the trace, debug a misfire.
WebFetch, WebSearch, MCP servers (GitHub / Slack / Calendar), and custom tools via the Claude Agent SDK. A mental model for picking the right tool — by I/O cost, blast radius, and confirmation needs.
Decompose a topic into search → read → verify → synthesize → self-review. The plan-then-execute pattern. Forcing citations. When the agent should ask vs. proceed.
Budgets (max steps, max tool calls, time caps). Idempotency & retries. Hooks (PreToolUse / PostToolUse) as guardrails. Observability — reading and sharing failed runs. Background vs. interactive. Where humans need to stay in the loop.
Choose one of five pre-scaffolded templates, customize it to your real life and data, run it, and demo in 90 seconds. Your agent. Your inbox or repo or calendar. Real automation.
In Module 6, every student picks a template, customizes the SKILL.md / agent definition to their real workflow, and demos it. These are the kinds of agents students typically walk out with.
Reads recent email, classifies by intent, drafts replies in your tone. You approve and send.
Scrapes 3 sources you choose, writes a morning brief at 7am. Runs on a schedule.
Reviews open PRs against your team's checklist — edge cases, tests, naming, performance — and posts comments.
Reads your week, drafts time-block proposals, flags conflicts before they become problems.
Your Module-4 agent, scaled to topics you actually care about. Daily, weekly, or on-demand briefs.
If you can describe the loop — goal, tools, when to stop — we'll help you scaffold it. Bring your data.
If you've finished Basics and Skills and you're thinking "what if Claude could just do this for me without me asking every time?" — that's the question this course answers.
Hands-on the whole time. One break in the middle. Show & tell at the end — most valuable part of the day.
Concept, demos, the loop diagram.
First working subagent.
Web, MCP, custom tools.
Multi-step, plan-then-execute.
Stretch, walk, share what you've built so far.
Budgets, hooks, observability.
Pick a template, customize, ship.
90-second demos, "what went weird" discussion, cheat sheet.
Yes. We move fast. If you can already write a SKILL.md and direct Claude through plain-English prompts, you'll keep up.
A skill is how Claude behaves on one task. An agent is a loop that picks tasks, picks tools, observes, and decides what to do next — until the goal is met or a budget is hit.
A laptop with Claude Code (set up in Basics), a Claude Pro account (monthly — agents make many tool calls), and real data: a folder, inbox, repo, or calendar you'd actually like automated.
Some — mostly small Python or TS functions for custom tools. Most of the workshop is YAML/Markdown agent definitions and reading traces. The Agent SDK is one tool among many.
Yes — we cover this in Module 5. Tool allow-lists, hooks, and dry-run flags are first-class. You always know what an agent can do before you run it.
By default: a subagent file in your project that runs on demand or on a schedule. We don't deploy to servers — that's a separate course. The agents you build run wherever you run Claude Code.
Small groups, max 8 students. Includes the workshop, take-home cheat sheet, agent templates, and the agents you build.
You'll leave with agents running on your machine, on your data — and the muscle memory to design the next ones yourself.
🔔 Notify me when availableReleased to Basics + Skills alumni first.