Hone discusses AI decision-making safety (Part 3)

This week. Now, the obvious worry is that you're giving an AI access to sensitive company systems and letting it run for months with a lot of leeway. Hone's answer is guardrails. Before it takes on any responsibility, an engine's decisions get simulated first. Anything sensitive needs a human to approve it. And everything gets logged so somebody can go back and look. Hone's second blog post goes deep on that simulation piece. It opens with Apollo 13, where engineers on the ground tested their fix in a simulator before they sent the instructions up to the astronauts. Hone's point is that today's AI agents mostly don't do that. They make a change and find out in production whether it broke something. The thing is, Hone's engines change themselves. They rewrite their own tools, their code, their memories, all as they learn any one of those changes can break something that's supposed to work. So before an engine locks a change, it replays it against past situations to see what else it might mess up. Their example goes like this. Imagine a manager named George tells the engine, check with me before you order from a new supplier. So the engine writes itself a rule. Always ask George first. But when they replay that rule against past situations, it turns out it breaks something. There was a case where a production line was down and somebody said, order the bearings now. And with the new rule, the engine would have sat there waiting on George instead. So the rule gets fixed too. Always ask George. Unless a line is down and it only goes live once it passes the old cases too. The hard part, Hone says, is that you can't rewind a company. You can't roll your salesforce back two weeks. So they built an agent they call the director that plays the role of the outside world during a replay, deciding when a person writes back or when a shipment notice shows up. Other agents stand in for each software tool. And Hone says those fake tools got the real facts right about 90% of the time in tests. They also forward the clock so weeks of work can replay in minutes. Hong claims that across 125 Test scenarios, replaying a change first cut the share of changes that broke something from 33% down to 5%. That is their own internal test, of course, and nobody outside has checked that yet. The founder, Stefan, is Austrian, and he's been starting things since high school. Back in Vienna, he and a partner launched a startup called.