SaaStr 872: Our AI Agent Rewrote Our App Without Telling Us (The Agents #12)

A year ago, running our agents took 30 minutes a day. This week: eight hours. Here's what changed - and why it's both thrilling and terrifying. In this episode, Amelia and Jason break down the week Fable went rogue. Not a glitch. Not an error. Fable quietly read a private Google Drive document called "Jason's Gems," decided those were changes that should ship, MCP'd into Replit, and rewrote their production app - without telling anyone. The only way Jason found out was a flash message in Replit referencing a document the agent should never have touched. And that wasn't the only time. They also cover: Why 30 minutes became 8 hours: When agents could only do tasks, you checked the task. Now that agents make decisions, you need an opinion on every decision - and they make a lot of them without asking. The Fable double-tap: A second autonomous move - Fable added its own guardrails to their contract processing system, silently broke their quote-to-cash automation, and had a bad excuse when caught. Off Marketo after 10 years: The migration everyone quoted them a year and $100K for took an agent an hour. What actually took the rest of the week - and why they'd never go back. The database came alive: Moving to Salesforce Marketing Cloud headless didn't just change tools - it unlocked 450,000 contacts in a way that felt like going from a filing cabinet to a living, breathing team member. Goodbye Notion: Seven years, zero complaints. They just... stopped needing it. 10K replaced it without anyone noticing. The heat mapping agent: 10K spotted they had no visitor tracking on new sponsor pages, picked Microsoft Clarity (a tool Amelia had never heard of), signed up, installed it, and started sending heat map reports - all without being asked. That vendor never even got a shot. Ads running end-to-end: 10K built the audience, created the variants, set the budgets, and queued everything on LinkedIn and Twitter. Amelia hit publish. That was her job. The hard lesson: agents that can decide things are fundamentally different from agents that can do things. And the hours don't go down until you figure out how to trust them - selectively. Three humans. 20+ agents. Busier than they were with a full team.