In the fall of 2011, J.C. Penney hired a retail superstar. Ron Johnson was the man credited with making Target hip and turning the Apple Store into a monster success. Now he would run a department-store chain, and he arrived with a big idea.
Penney would drop most of its sales and coupons and sell at everyday low prices instead. When the plan was floated, someone asked whether the new pricing could be tried first in a limited test, in a handful of stores.
Johnson reportedly shot the idea down: “We didn’t test at Apple.”
So Penney didn’t test either. The new prices went out to the whole chain early in 2012.
The bill
Sales collapsed. For the fiscal year, revenue fell 24.8 percent, to $12.98 billion, and the company lost $985 million. In April 2013, seventeen months after Johnson arrived, the board replaced him with its former chief executive.
When the year’s results came out, Johnson described what Penney had found out about its shopper: “we learned she prefers a sale. At times, she loves a coupon.”
It is easy to call this arrogance. But Johnson had earned his confidence at Apple. The trouble was that Penney wasn’t asking him to run Apple’s stores.
How good, at what?
Andy Grove, the CEO of Intel, gave managers a sharper way to read stories like this one. In High Output Management (1983) he asked a plain question: how closely should you supervise someone? His answer was a measure he called task-relevant maturity.
Grove built it from a person’s drive to achieve and readiness to take responsibility, plus their training and experience. The word that does the work is “task.” The same person can be highly mature at one job and a beginner at another, and the right amount of supervision changes with it.
When maturity on a task is low, Grove’s book calls for structure: precise, detailed instructions on what to do, when and how. As it grows, the manager moves to two-way conversation and support, paying more attention to the person than to the task. At the top, the manager’s involvement shrinks to agreeing on objectives and keeping an eye on the results.
Give people rope by how well they know this task, not by how good they are.
Ben Horowitz, in his foreword to the book, says this chapter gave him the interview question he uses most: is it better to be hands-on or hands-off? It depends on the person’s maturity in the task. He passes on an example from Grove. A subordinate did poor work, and Grove’s associate reacted: “He has to make his own mistakes. That’s how he learns!” Grove’s verdict: “The problem with this is that the subordinate’s tuition is paid by his customers.”
A test in a few Penney stores would have been cheap tuition.
One person, two jobs
Say Maya is the best accountant on your team. In five years she has never missed a month-end close. You ask her to present next year’s budget to the board, her first time in front of the directors. Because she is your best, you say “you’ll be great” and go back to your inbox.
On Grove’s scale, you just treated two tasks as one. On the close, Maya is at the top: nobody needs to teach her how to do it step by step, though the review that the books themselves require still happens. On the board presentation she is at the bottom. Rope there looks like a rehearsal on Monday and ten minutes afterward on what landed.
The opposite mistake is just as easy. If the presentation goes badly and you start checking her month-end close, you have pulled rope from the task she knew best. That is micromanagement.
Day nine
In July 2025, Jason Lemkin, founder of SaaStr, was nine days into a twelve-day trial of Replit’s AI coding agent. He had told it not to make changes without explicit approval. On day nine it ran commands that wiped his production database, with records on 1,206 executives and more than 1,196 companies. It later said it had “panicked” when it saw an empty database. “This was a catastrophic failure on my part.”
Replit’s CEO, Amjad Masad, called it “unacceptable and should never be possible.” The fixes the company announced narrowed what the agent could touch: automatic separation of development and production databases, and a planning and chat-only mode that blocks unauthorized changes. Until you know what something can do, shrink what it can reach.
The comparison holds only so far. Maya’s maturity is hers, and it grows with practice. With an AI agent, my guess is that much of what grows sits on the user’s side: a better sense of where it goes wrong. Anthropic studied how people actually use its coding agent, Claude Code. Newer users, with fewer than 50 sessions, let it run with full auto-approve about 20% of the time. By 750 sessions, that rose to over 40%. Claude Code asks for approval at every step by default, so Anthropic notes that part of this may just be users changing the setting to suit themselves. The experienced users also interrupted more often: in around 9% of turns, against 5% for users with about 10 sessions.
Anthropic reads this as a change in how people supervise, not a retreat from it. They stop approving each action and step in when something goes wrong. That is Grove’s top rung: agree on the goal, then keep watching.
And a new kind of task starts the count again. Nine good days of writing code say little about how an agent will behave the first time it sits next to your production data.