A Different Job Filter
If I were looking for a job today, I would ask a question that barely existed a few years ago: how much are the people here actually working with agents?
Not whether the company has an AI strategy. Not whether it bought everyone a coding assistant. I would want to know if the strongest people are agent maxxing: running several agents, delegating real work, building their own loops, comparing workflows, and spending enough tokens that nobody is nervously watching the meter.
That would matter more to me than a lot of the usual signals. You are partly choosing a company, but you are also choosing the environment in which you will learn how work gets done next.
You Need Repetitions
Working well with agents is not something you learn by reading a prompt guide or watching someone finish a polished demo. You learn it through repetitions. You give an agent a task, watch it misunderstand the shape of the problem, figure out which context was missing, change the way you delegate, and try again.
Over time you start noticing where agents need freedom and where they need constraints. You learn when to keep one thread going, when to start over, when to run work in parallel, and when the output looks plausible but has not touched the real problem. None of this becomes instinct after ten prompts.
The people getting unusually good at agents are not necessarily the ones with the best theories. They are the ones compressing hundreds of these loops into their normal work.
Unlimited Tokens Change Behavior
This is why unlimited tokens are not just a nice employee benefit. They change how willing you are to practice.
When usage is tightly rationed, every interaction starts to feel precious. People save agents for work that looks important enough. They avoid parallel attempts. They accept mediocre output because another pass feels wasteful. They use the tool carefully when the useful behavior is often to use it aggressively, inspect what happened, and run the loop again.
The cost is not only fewer tokens. It is fewer experiments. Someone with effectively unlimited access can try three approaches, give the same task to different agents, let one investigate while another implements, and throw away work without treating it as a failure. That person will develop judgment faster because they see more outcomes.