Why Trust Is the Bottleneck for Agents — Jean-Denis Greze, Town | AI Engineer World's Fair 2026
B · Useful, limited Short AI Engineer World's Fair interview with Jean-Denis Greze, former Plaid CTO and founder of Town, giving a concrete permission-escalation pattern for agents; useful practice, but anecdotal and partly a product pitch.
Summary based on the full transcript·Original: about 16 minutes to watch
In short
A WorkOS interview recorded at the AI Engineer World's Fair 2026 with Jean-Denis Greze, ex-Dropbox and former Plaid CTO, now building Town, an email and calendar assistant for non-technical people. His argument: trust, not model capability, limits agents in production. Town starts with conservative, draft-only automations and earns permission for bigger jobs over time.
Key numbers
- 10prior approved actions Town uses before asking a user to let an automation run without approval
- 10 to 15reply templates that cover most emails for roles such as recruiters and salespeople, in Greze's estimate
As reported by the source.
English·Video·opens youtube.com
What it reports
- Town pivoted from AI for tax preparers after noticing their email and calendar overhead; the prototype took two weeks and got 20 DAUs within days.
- Rule: never send an email that makes the user look stupid; one bad action on day three loses the user.
- By default no external write actions: the agent puts drafts in the outbox rather than sending.
- Permission escalation: after about 10 approved actions of the same kind, Town asks whether to stop asking. If users edited or discarded drafts, it does not ask. The agent cannot change its own permissions.
- Horizontal use cases: replying to common emails (10 to 15 templates for many roles), scheduling, meeting preparation, to-do tracking.
- Feedback built into failures: an in-product bug button and opt-in capture of unsupported requests (for example travel planning) feed the roadmap.
- Multi-user vision: agents in separate data silos sharing derived signals, such as a relationship-strength score for a prospect, without exposing raw email.
How they got it
Founder interview; no data beyond the early usage anecdote.
Limits
- Town had launched only a month earlier.
- Promotional context; no measured trust or retention figures.
- The multi-agent privacy idea is speculative.
Why it matters for companies
The approval ladder is a pattern internal agent teams can copy directly: draft-only by default, track the human edit rate per action type, offer autonomy only where acceptance is consistent, and keep permission changes in human hands. It treats trust as something measured per workflow.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- Lecture