Agent UX: Designing Trust
How much should you hand over to an AI agent — approval queues, visible reasoning, interruption budgets. Working hypotheses and observations on designing trust.
This note is still being written.
Problem
Give an agent “do anything” permissions and users feel anxiety before convenience. So how much autonomy actually builds trust? Working hypotheses from data observed across three products in the first half of 2026.
Observation: the approval budget
Users tolerate roughly 5 confirmation dialogs per session on average. From the sixth onward, they split into two camps:
- Approve-everything (“fine, whatever” mode) — dangerous
- Reject-everything (“I’ll do it myself” mode) — churn
So the number of prompts is a design resource: once it’s spent, it’s gone.
Pattern: three levels of trust
- Visibility — what the agent will do next is always listed (a queue)
- Reversibility — every action has an “undo”
- Staged delegation — permissions expand only through accumulated small wins
Does showing reasoning increase trust?
Counterintuitive finding: detailed reasoning traces lowered novices’ trust. Noise piles onto the “I don’t know what I don’t know” state. One-line summary + collapsed details is the current optimum.
Next experiment
If we visualize the approval budget (“2 confirmations left”), does approve-everything mode decrease? In planning.
This is a living document — I’ll update it as observations accumulate.
