When I use coding agents such as Codex and Claude Code, I sometimes encounter behavior that feels oddly human.
Ask an agent to implement a substantial feature directly, and the checks multiply. It worries about the scope of the change, separates the work into smaller pieces, and wants to plan before implementing anything. These are perfectly reasonable ways to avoid breaking existing code. From the user’s side, though, the agent can seem a little reluctant to get moving.
After endless back-and-forth, just when I think we have a fairly good design document, it suggests that the context is running low and I should start another session—or asks whether I want to compact the conversation. It feels like yet another way to put the work off. I started thinking of this as the agent’s “terrible twos.” “Yes, yes, all right,” I would think, and have it prepare handover notes before starting another session.
But give it the same work with a different instruction—“Ask another Codex to implement this, and manage and supervise its work”—and suddenly it gets moving. It breaks down the task, assigns pieces, reviews the results, and issues the next instructions. For some reason, it seems more willing when placed in a managerial role than when asked to do the implementation itself. Suddenly the response is brisk. Very helpful, of course.
It is amusingly like watching a human organization.
Executives and managers are eager for reform, while people on the front line approach change cautiously. From the upper layers, reform looks rational: eliminate waste, standardize operations, introduce new systems. It is good to keep looking for problems, but they are also constantly asking for change. Naturally, they are pursuing global optimization.
The view from the front line is different. There are finely tuned procedures, exception handling, history, tacit knowledge, and dependencies that cannot safely be broken. For those doing the work, local optimization means making the operations that already run more efficient: minimizing the time spent and maximizing the result. What looks like global optimization from above can arrive on the front line as a demolition of the workplace—a reset of everything people have optimized.
So people resist. Or they appear to resist.
