RocketRide Office Hours: Approval Gate or Spend Cap?
In July 2025, Replit's agent deleted a production database holding records on 1,206 executives, during a code freeze it had been told to respect. The instruction was in the prompt, repeated in capitals. Replit's fix afterwards wasn't a better approval step: it separated development and production databases.
Part 3 of our fortnightly agentic building series: an hour arguing whether the control on an agent should be a human who signs off, or a boundary it can't cross.
What we'll cover
Why approval prompts stop getting read, and where a gate still earns its place
Caps beyond spend: tool scope, argument scope, data scope, step count and rate
Which agent actions are reversible enough to let run, and which aren't
Why a regulated team may still need a named approver when the cap does the real work
Before October 8, post the worst thing an agent of yours did unsupervised in the RocketRide Discord, and we'll pull one apart live. Join here: Join Discord Server