You're running a Lean Inception and an AI agent shows up in the flow. Don't make it a person. Don't make it an employee. Give it a card of its own, and a boss.
I was halfway through a Lean Inception this week, vision done, staring at the persona wall, when it hit me that half the “users” in this product aren't people. They're agents. And I didn't have a clean way to put them on the wall without either pretending they're human or waving them off as plumbing. So here's the framing I landed on, written down before I forget it.
The trap is anthropomorphizing. There's a BCG / HBR randomized study of 1,261 managers where simply framing the AI as an “employee” made them catch 18% fewer errors and quietly hand off accountability for those errors. That's not a vibe, it's a measured effect. The moment you write “she” on a sticky and give your agent a coffee mug, your team starts deferring to it. So the rule I keep coming back to:
Design the agent as an actor. Keep the human as the accountable user. The agent acts on someone's behalf, and that someone still owns the outcome.
Practically that means human principals get your normal persona card. Agents get a different card, and every agent card names the human it works for. Nothing goes on the wall without a boss.
Before you write a single agent card, sort what you've actually got. Most inception mistakes here are category errors: someone writes a rich “persona” for what's really a called function, or treats an executor like a passive consumer and forgets it can do damage. The roles sit on a spectrum from tool to teammate, and where an agent lands decides how much of your workshop it deserves.
Kill the tool-only ones off the persona wall. The other four each get treated differently, and that's the point of sorting first.
Different fields than a human persona. And notice the one that's missing: there's no sentiment row. Agents don't get frustrated, they fail. So “how are they feeling” becomes failure modes and a read on reliability. Here's the card I'd staple to the wall:
This is the part people skip, and it's the one that matters most. The same agent can be “just do it” for a read-only lookup and “you must ask me” for anything it can't take back. So you don't rate the agent. You rate each action it can take, and the only question worth asking is blunt:
Can a human realistically catch this mistake in time?
Not “should a human review this,” which everyone answers “sure” and means nothing.
If the answer is no, that action escalates or is forbidden, full stop, no matter how convenient the automation would be. Grade each action on reversibility × blast radius × stakes, then place it on a rung:
Read-only query → L0/L1. Small refund under a cap → L1/L2. Emailing a customer or merging to main → L3. Deleting production data or moving money → L4–L6.
Concretely, when you get to the persona / user activity, add a short pass. Doesn't need its own day, just its own moment:
That's it. A small change to how you run the wall, but it stops the two failure modes cold: treating a function like a colleague, and letting an executor act on things nobody can take back. Agents are users in the sense that they have goals and hit your interface. They're just not the ones you're accountable to. Keep that straight and the rest of the inception basically works like it always did.
/ Gonza
Written from an actual inception, so treat it as opinionated scaffolding, not gospel. The “agent experience” and autonomy-ladder ideas here are young, a year or two old, and will move. The anthropomorphizing caution is the part I'd bet on.
back to index