Draft — replace with a hook about the first time a user said "wait, what's it doing right now?"
The problem agent UX has to solve
- The new contract between user and product when an agent takes actions
- Why the trust gap is wider than the automation gap
- What "handing over the keys" actually feels like to a first-time user
Three modes, one question
- Alerts only — user stays in full control
- Drafted messages — user reviews before send
- Fully automated — user trusts the agent end-to-end
- The single question all three are really answering
What earns trust
- Visibility into what the agent is about to do, not just what it did
- Reversibility on every action
- Honest failure messages — "I didn't try this because…"
- Pacing — slow when stakes are high, fast when they aren't
What loses trust fast
- Silent retries
- Over-confident summaries when the agent didn't actually understand
- Surprise actions that don't show up in any review surface
Open questions I'm still sitting with
- How do you let users grow the agent's permissions over time without forcing them to read a settings page?
- Can a trust score be visible without being anxiety-inducing?
- What's the agent UX equivalent of a seatbelt — present, easy, almost invisible?