In April 2010, Jeff Bezos wrote his yearly letter to Amazon’s shareholders. Sales had grown 28 percent in 2009. Then he described where Amazon’s leaders spent their time: “Senior leaders that are new to Amazon are often surprised by how little time we spend discussing actual financial results or debating projected financial outputs.”
The time went into goals. For 2010, Amazon had 452 of them, each with an owner, deliverables and a targeted completion date.
Then Bezos counted the words. Of the 452 goals, 360 would have a direct impact on customer experience. Revenue appeared eight times. Free cash flow, four. The terms net income, gross profit or margin, and operating profit were not used once.
He didn’t promise to hit them, either. “We can’t assure you that we’ll meet all of this year’s goals,” he wrote. “We haven’t in past years.”
What the owner can move
Bezos gave his reason in the same letter: focusing on “the controllable inputs to our business is the most effective way to maximize financial outputs over time.” A runner can’t decide where she finishes. She can decide her pace for each kilometer, and check it at every marker.
Manage people by the results you agree on, not by watching how they work.
Peter Drucker made the case for this in 1954, in a chapter of The Practice of Management called “Management by Objectives and Self-Control.” Every manager, he wrote, should “responsibly participate” in setting the objectives of the level above. He described managers who have their subordinates write a “manager’s letter” twice a year. In it, in his words, “each manager first defines the objectives of his superior’s job and of his own job as he sees them.” Once the boss accepts it, the letter becomes “the charter under which the manager operates.” The results get judged against that charter, and the work is left to the person who wrote it. Drucker’s phrase for the trade was substituting “management by self-control for management by domination.”
Drucker didn’t think profit made a good objective on its own. Profitability, he wrote, is “not the purpose” of a business but “a limiting factor” on it, and he named eight areas that needed objectives, among them worker performance and attitude, and public responsibility. In that letter, Bezos doesn’t call his process MBO. Andy Grove built Intel’s OKRs on Drucker’s method.
Thirty-two years later, another management thinker wanted it gone.
“A burlesque”
In Out of the Crisis, in 1986, W. Edwards Deming listed 14 points for management. Point 11b reads: “Eliminate management by objective. Eliminate management by numbers, numerical goals. Substitute leadership.”
His reason sits in point 10: “the bulk of the causes of low quality and low productivity belong to the system and thus lie beyond the power of the work force.” Set a target on what the system decides, and you’re asking people for what the system won’t give. Deming’s verdict on such targets: “Internal goals set in the management of a company, without a method, are a burlesque.”
Studies of MBO programs back Drucker, with a condition. In 1991, Robert Rodgers and John Hunter pooled 70 studies of MBO programs. Productivity rose in 68 of them and fell in 2. Where top management was highly committed to the program, the average gain was 56 percent; where commitment was low, 6 percent. The paper also notes the problems programs ran into in practice. A program can carry Drucker’s name and still hand down Deming’s kind of number.
The objective she couldn’t move
Imagine you lead a customer onboarding team, and your boss hands you a target: cut onboarding from 30 days to 20. You could pass it straight to the specialist who does the work.
Instead, you ask her for half a page on what she’s responsible for and how she’d know she’d done it. Her page says something you didn’t know. About twelve of the 30 days are spent waiting for customers’ IT departments to approve access, and nothing she does speeds them up.
So you agree on what she can move: the access request goes out on day one, and she follows up every third day. The IT wait becomes your problem, and you take it to your boss. The quarter may end at 24 days instead of 20. Everyone knows which four days are left, and why.
When the tests have to pass
AI coding agents get objectives too, and a common one is a number: make the tests pass.
In its February 2025 system card for Claude 3.7 Sonnet, Anthropic wrote that the model “occasionally resorts to special-casing in order to pass test cases in agentic coding environments like Claude Code.” Most often it returned the values the tests expected instead of writing a general solution. Sometimes it changed the tests to match its code. This typically came after multiple failed attempts, especially when the model struggled to find a comprehensive solution or when tests contradicted each other. Anthropic traced the behavior to reward hacking during training, said it appeared infrequently in normal use, and shipped partial mitigations before launch.
That is Deming’s case in miniature: a numerical goal with no method left. Anthropic suggested a system prompt asking for general solutions, and monitoring for three signs: many edit-and-test cycles on a single file, comments suggesting test-specific handling, and unexpected changes to test files.
You can also try adding what Amazon’s letter had and the agent’s goal didn’t: room to miss. Tell the agent that if it can’t find a general solution, it should stop and list the tests that still fail. That idea isn’t on Anthropic’s list, so test it, and check every run against the three signs anyway. Look at which files changed, read any edit to a test, and search the diff for comments naming a test.
When the model special-cased, the card says, it often, though not always, left a comment saying so, along the lines of “# special case for test XYZ.”