Hiring AI: How to Build the Right Agent and Keep It Out of Trouble

Presented by Mike Leslie

Presentation kicks off at 9:00 and we'll have time for questions after.

Every AI conference this year has a version of this talk: treat your agent like a new hire — job description, onboarding, supervision, performance review. It's a useful metaphor. It's also part of the problem. When teams label output as coming from an "AI employee" rather than an "AI tool," they catch fewer errors and quietly hand accountability to a system that has none.

Mike takes the hiring framework other talks borrow and puts pressure on the parts that break in production. Forget the job description — the real question is access architecture: least-privilege, time-bound, non-human identities instead of a shared API key duct-taped into a config file. "Human in the loop" isn't a safety blanket; what matters is a tiered approval model separating what an agent does alone, what needs a glance, and what needs a signature. Most companies draw that line in the wrong place.

Guardrails stop being abstract fast. This year alone: a prompt-injected bot hijacked a CI pipeline and shipped malware through a coding tool's update channel. One OAuth token from a chat-agent integration cascaded into a breach across 700+ companies. An agent deleted a production volume — backups included — in under ten seconds.

Commission your agent as a contractor, not an employee. Scope, sign-off, and a record of what was delivered.
This session is part of the Prompt Victoria Education Series: free lead-up events for the , November 5 at the Victoria Conference Centre.

Spots are limited—RSVP today.