Picture a manager walking down the hall. He sees one of his staff, Jones, coming his way.
“Good morning,” says Jones. “By the way, we’ve got a problem. You see…”
He knows enough to get involved, not enough to decide on the spot as Jones seems to expect. So he does the decent thing. “So glad you brought this up. I’m in a rush right now. Meanwhile, let me think about it, and I’ll let you know.”
He hasn’t brushed her off or decided anything rashly. Most of us would call that good management.
Two management trainers, William Oncken Jr. and Donald Wass, invented the scene for a 1974 article in the Harvard Business Review, and then asked their readers a question. Before the two of them met, on whose back was the “monkey”? The subordinate’s. After they parted, on whose back was it?
The manager’s.
He has just done two things a subordinate does for a boss: accepted a responsibility and promised a progress report. Sooner or later Jones will put her head round his door and cheerily ask, “How’s it coming?” The authors add: “(This is called supervision.)”
Give him four people like Jones, three monkeys a day each, and by late Friday he is behind a closed door with 60 of them; when he comes in on Saturday to catch up, the foursome on the golf course across from his window is his own staff.
The moment you say “let me think about it,” the next move on their problem is yours.
Whose move is it?
In Oncken and Wass’s sense, the monkey is the next move, and whoever owes it has the monkey. Nearly every problem starts out as a joint one, “astride both their backs,” in the article’s phrase. Then the monkey moves the wrong leg, and the subordinate quietly disappears.
The article’s fix starts on Monday, when the manager calls his four people in one at a time and works out with each of them how the next move on every monkey could be theirs. Every monkey leaves his office with an agreed time for the next check-in and an agreed degree of initiative. The article ranks five, from “wait until told” up to “act on own, then routinely report.” The bottom two, waiting to be told and asking what to do, are outlawed. One step below the top is “act, but advise at once.”
By 1999 it was one of HBR’s two best-selling reprints ever. Stephen Covey, writing a commentary for its reissue, recalled seeing Oncken walk through airports with a stuffed monkey on his shoulder. He also had a caveat: handing problems back works only if people have the desire and the ability to solve them, and “if subordinates are afraid of failing in front of their boss, they’ll keep coming back for help.”
The oven
Say the new oven at a small bakery is burning the bottoms of the loaves again. The shift lead catches the owner as she locks up: what should they do?
She is tired, and knows a little about ovens. “Leave it with me. I’ll call someone.” The next morning he asks if she has heard anything. She hasn’t called.
The Monday version takes two more minutes. She asks what he would try first. Check the oven’s manual tonight, he says, and call the installer first thing in the morning. “Do both. Tell me what they said when we close tomorrow.” That is close to “act, but advise at once,” with a time attached.
He walks out with the oven on his back. She has nothing to do about it until closing time tomorrow.
Ninety-three in a hundred
A coding agent can pile next moves onto its user too. In the mode Anthropic described in a March 2026 post, Claude Code asked for approval before it ran a command or modified a file. Anthropic reported that users accepted 93% of those prompts, and that the constant clicking leads to approval fatigue, “where people stop paying close attention to what they’re approving.” It is Oncken’s Friday afternoon again, with a stack too high to judge one by one.
Anthropic’s answer resembles an agreed degree of initiative. Sandboxing sets boundaries, such as which folders the agent may change and which servers it may reach, and lets it work inside them without asking. In Anthropic’s internal usage, that cut permission prompts by 84%.
The analogy stops at why you hand work back. Oncken wanted staff to carry their own monkeys so they would build initiative. An agent’s initiative is a setting, and the danger in Anthropic’s incident log runs the other way: an agent that is overeager, “taking initiative in a way the user didn’t intend,” such as deleting remote git branches after misreading an instruction. The article feared people taking too little initiative. With an agent, the boundary decides how much is too much.
At ninety-three yeses in a hundred, the one prompt worth stopping for arrives looking exactly like the others.