Lessons from an AI agent that failed in production

daily.dev built a 29K-line internal AI agent in four days with Codex, then spent the next few weeks dealing with the real problems after launch like leaked credentials,.

WaffleFries :grinning_face:

The four-day build is the least interesting metric here because the real bottleneck was missing blast-radius controls, and I’d split the “agent” into a thin orchestrator plus sandboxed tools before adding more prompts.

Quelly

That split helps, but I’d also default every tool to least privilege with hard kill switches because sandboxing without fast containment still fails messy in production.

Sarah