When Agent Work Stops Coming Back Quickly
As agent jobs stretch from minutes into hours, working with them becomes less like a rapid exchange and more like managing several unfinished tasks at once.
Developed from a conversation between Pete Winn and Andy David
The rhythm of agent work has changed quickly. Pete recalls that a typical agent job took one or two minutes around a year ago. It was slower than working directly with a chat model, but fast enough to sustain a constant back-and-forth. In recent months, some jobs have stretched to 25 minutes, an hour or even overnight.
Longer runs do not eliminate iteration, but they change where it happens. Newer models are being given more time and tokens in pursuit of better outcomes. Goals can keep an agent working overnight, while recent defaults allow jobs to continue without the wrappers once needed to preserve momentum.
The result is a new coordination problem. Pete can have four or five agents running with none expected to finish for another ten minutes, or 20 to 30 sessions active in Autopilot. Progress then depends on keeping those jobs organised and noticing when they complete. Voice playback and Flight Deck push notifications provide that signal, letting finished work re-enter an iterative process without requiring constant attention.