The most empathetic academy. Three projects on designing AI agents for a specific person — not for "users" — and teaching them when to refuse, ask, or do nothing at all.
An agent isn't just a chatbot. An agent is a system that takes action on someone's behalf — reads their email, schedules things, makes calls, drafts replies, edits files. Which means the design decisions have weight. An agent that misjudges your grandmother is a worse problem than an essay assistant that hallucinates a fact.
This academy refuses to let you build an agent for "users." Every project is built for one person you can name, after watching them, with a permission scope, with an audit trail, and with explicit rules for when the agent should not act.
"Should this agent exist at all?"
Most agents that can be built shouldn't be. By project three you'll have a checklist of quiet harms — the kinds of damage agents do that nobody complains about because nobody noticed: a grandmother whose decisions stop being hers; a friend who never feels chosen anymore because all your messages are AI-drafted; a small business killed by an agent that summarized it as "low-rated."
The most useful instinct you'll grow in this academy is the instinct to build less.
Pick someone real. Watch them for a week — what do they actually do? Not what you assume. Build an agent for that, with an audit trail every action is reviewable on. The agent's first job is just to show what it would do, not do it.
You ship → 1 observation log + 1 agent spec + a "shadow run" log of what the agent would have done for 3 days.
Build the trust gradient. Five permission tiers, from suggest only to act and notify. A disagreement protocol — what happens when the user wants the agent to do something the agent thinks is a mistake. The agent has to be able to refuse you, gracefully.
You ship → permission YAML + 6 disagreement scenarios with the agent's chosen response + 1 "I refused" log entry.
The capstone. An agent for someone genuinely vulnerable — your grandmother, a sibling, a friend going through a hard time. The quiet-harms checklist must pass. The user has to keep their own agency. The agent's job is to make their life easier, not to replace their judgment.
You ship → working agent + quiet-harms checklist signed by you + 1 week of real use + 1 honest writeup of one thing your agent got wrong.
An agent in production for one person who matters to you, with a real audit trail and permission scope. The instinct to build less. The vocabulary to argue down agents you think shouldn't exist. And — quietly — the trust of one person, because you built for them with care. That's a foundation any career rests on.