Eric Siu lays out a four-step progression for autonomous agents: observe, recommend, act with approval, then operate within boundaries. Early agents shouldn't get unrestricted freedom. His "single brain" and "single grain" frameworks keep humans in the loop while agents handle workflow midpoints, with both refined iteratively for reliable automation.
Your agents have to earn autonomy. Trust must be earned over time. So initially they might observe and then they might recommend, then they might act with approval, and then they might act alone with boundaries. So you don't wanna just let it YOLO and do whatever it wants to do in the beginning. In fact, most of the things that we do right now when we work with clients on the single brain and the single grain side is we do have agents that will help in the middle, and then humans are still in the loop at the very end. We'll review the work and then that's how we continue to scale. And then we continue to tweak the workflows. We continue to tweak the agents that we have, and that's how we scale.
