The trust ladder: how much do you trust your AI agents?
There is a lot of buzz about AI software factories at the moment. One of the big questions in this area is "how can I trust the code is good when no human has reviewed it?" There is an interview between Lauren Tan (aka Poteto) and Matt Pocock that has some great thinking on this:
During the interview Lauren uses the concept of a trust ladder. The basic idea is that you start small with only a few AI agents because you don't trust them yet. You carefully monitor the output and conversations and then correct course where you need to. As you need to correct less, you can give the agents more autonomy - because you can trust them more - this is how the agents climb up the trust ladder. Once your agents have gone up the trust ladder far enough, you can then scale up the number of agents.
I think a trust ladder is a really useful mental model when trying to create some sort of framework to build trust with your agents. In practical terms it means building guard rails into your AI workflows, so that the agents can correct themselves. I suspect most software engineers who use AI to write their code have intuitively thought about this in some way. But until now I haven't had a useful term that crystallises this thinking into a concept that I can talk about and share.