Intermediate AI: Hands-On AI Agents

At a glance

About this course

A hands-on working session in which professionals move from AI chat users to supervisors of autonomous AI agents. Delivered in two editions — Anthropic Claude (agents across every Claude surface) and Microsoft Copilot — participants work on their own laptops to install and operate an autonomous AI agent, write steering files that encode their working rules, put real folders under git version control using plain-English commands, and climb a graduated “autonomy ladder” — permission allowlists, sandboxes, and a machine-reviewed auto mode — that trades babysitting for engineered safeguards. They enforce hard rules with event hooks, practice plan-first supervision by editing build plans before any work executes, and finish with two measured capstones: reverse-engineering a text classifier from labeled data to 90%+ tested accuracy, and planning, building, versioning, and extending a working personal CRM application. Governance principles from Day 1 — least privilege, reversibility, containment, human oversight — are applied at every rung.

Instructor: Aaron Sparks, Founder, edje. Delivered in person, on-site at your office in Seattle and throughout the Pacific Northwest.

What you'll learn

Picks up where Intro to AI leaves off and moves you from chatting with AI to supervising an agent that does the work. Nine hands-on exercises build toward two extended capstones, with practice datasets provided (75 labeled customer inquiries and starter CRM contact data).

• One assistant, every surfaceChoosing chat, document workspace, desktop agent, or terminal agent based on audit-trail, measurement, autonomy, and building needs.
• Your agent, installedSetting up and operating an autonomous agent (Claude Code in the Claude edition; equivalent Copilot tooling in the Microsoft edition).
• Steering filesStanding rules (e.g. CLAUDE.md) for versioning, secrets hygiene, style, and approval boundaries.
• Git in plain EnglishCommit, diff, restore, branch, and worktree, so every agent change is reversible and auditable.
• The autonomy ladderAllowlists, sandboxes, and machine-reviewed auto mode, matched to task risk.
• HooksEvent hooks that enforce rules an agent cannot override.
• Plan before buildPlan mode and a structured requirements interview, so you edit the plan before any work runs.
• Choosing the right modelMatching model tiers, including long-running frontier models, to task complexity and duration.
• Capstone: the classifierDeriving a text classifier from labeled data and testing it to 90%+ accuracy on a held-out set.
• Capstone: your local CRMPlanning, building, versioning, and extending a working app by reviewing diffs, not code.

Logistics & what to bring

Frequently asked questions

Also in the catalog