↳ design systems that act on behalf of real people
Agent Lab is where kids stop thinking about AI as a chat window and start thinking about it as a thing that acts — on behalf of a specific person, with scope and pauses and permissions. The most empathetic academy, because every agent is for someone who has to live with the result.
↳ an agent is a program with its own goals and methods
Build your first agent simulator. Help grandma book a dentist appointment. Discover that agents are easier to reason about than they look.
↳ observation before assumption
Real observation, not personas. Interview someone. Watch them. Then design.
↳ narration is trust
Build an agent with a Silent mode and a Narrated mode. See which one you'd actually use — and why the quiet ones make you nervous.
↳ pausing is a feature, not a bug
Learn to design the pause points — the moments where an agent stops and asks before doing something irreversible.
↳ no agent should be allowed to do everything
Write a permission manifest for your agent. Draw a line around what it can and cannot touch. Learn the trust gradient.
↳ sycophancy is a failure mode, not a feature
Teach your agent to push back when it thinks you're wrong — and to do so without being a jerk.
↳ the worst harms never make headlines
Engagement traps, learned helplessness, optimization mismatch. The five-question audit that catches them all.
↳ if you can build for someone vulnerable, you can build for anyone
The Agent Lab graduation project. Build a full agent spec for Wen, 76. Every technique from every previous module, all at once.