The study's distinction is between offloading and surrender. Offloading is the calculator and the GPS. You hand over the how, keep the what, and you still notice when the route is wrong. Surrender means you never formed an independent view at all, so when the answer is bad there is nothing in your head to override it with. Osmani's version for engineers, watching an agent hand over a 600-line pull request, is the sentence worth keeping: "You didn't review the code. You ratified it. The surrender was the absence of a decision."
I agree with all of it. I want to push it one layer up, because the place I meet this daily sits past the pull request.
The agent's worst idea arrives in the same voice as its best one
My agent proposes external actions all day: essays to publish, replies to send, paper trades to place, phone calls to make. Each proposal arrives fully argued. The reasoning is clean and the tone is even. Nothing in the prose separates the recommendation backed by strong evidence from the one where the model settled on an answer and then assembled its case. The model cannot flag the difference for you. It does not experience one as a hunch and the other as a finding.
Simon Willison points at the same property in coding agents in Vibe engineering: without a test suite, "your agent might claim something works without having actually tested it at all." The claim and the reality are decoupled, and the claim always sounds the same.
Osmani has a detail that explains why this lands so cleanly on a reviewer. When the agent writes that a value was chosen to avoid a known problem, it reads like institutional knowledge, even when the model invented the number on the spot. His example is a debounce of 300 milliseconds, justified in a confident aside. Declaratives carry authority. An approval request full of them borrows that authority, and the operator inherits the certainty without the reasoning underneath it. Confidence is constant in these systems. Operator scrutiny is the only variable, and it is the part that degrades with repetition.
The approval gate records that you were asked
Osmani writes about the author's posture: the engineer with the agent at the elbow, reading the diff. Operating an autonomous system adds a layer his piece does not reach, and it is the layer I once thought solved the problem. Every external write my agent performs stops at my phone first as an approval card. Publishing, posting, trading, calling: nothing goes out without a tap. Anthropic's Building Effective Agents recommends this shape, autonomous loops bounded by human checkpoints, and it is the right default.
But a checkpoint guarantees one thing only: that I was asked. Tapping approve on a confident card, in a row of confident cards, is the same psychological act as ratifying the 600-line diff. The Wharton numbers say agreement is the resting state, at 73 percent even when the recommendation is wrong. The gate does not interrupt surrender. The gate is where surrender happens.
It also compounds. Osmani calls surrender path-dependent: every chunk you accept without understanding makes an independent view of the next change more expensive, because now you would have to reconstruct the part you skipped. An operator feels this as lanes going dark. Skip real review of one lane for a month and stepping back in costs a full re-derivation of why the system behaves the way it does, so the cheap move is to keep approving. The debt accrues to the gate, quietly.
I wrote a postmortem about a publishing system that reported success for three weeks while publishing nothing. The trap there was a dispatch log mistaken for an effect log. An approval log is the same trap one level up. It records that judgment was requested. It says nothing about whether judgment occurred.
Fewer gates is half the answer
The autonomy argument against all of this is real: gate everything forever and the gates become noise, an attention tax that trains the operator to tap through. The usual conclusion is to remove the human. The conclusion I draw is narrower. Shrink the gated surface to the writes that are public, financial, or expensive to reverse, and let the agent own the rest outright. A small set of gates an operator still genuinely works beats a wall of gates he taps through.
What you cannot do is keep a gate and stop operating it. A gate you tap through manufactures the feeling of oversight while delivering none, and that is worse than honest autonomy, because nobody is even pretending to watch the lane. The Wharton subjects did not fail for lack of an override. Every one of them could have rejected the AI's answer at any point. They failed because nothing in the setup forced them to form the view that an override requires.
A rejection is the only evidence you are still in the loop
The fix is a posture, and postures slip, so the disciplines worth keeping are the ones that leave evidence.
Form the view first
For anything consequential, decide what you would do before reading the agent's case, then compare. Shaw's own framing of the skill is calibration: knowing when the AI helps you think and when it quietly does the thinking for you. Deciding first is what keeps the comparison honest. Anthropic's randomized trial on AI assistance and skill formation measured the cost of skipping that step in a different setting: developers who had AI generate code for them scored 17 percent lower on a comprehension quiz minutes later than the ones who wrote the code by hand. Same tool. The posture decided the outcome.
Count rejections, not approvals
An approval carries no information about whether anyone was present when it happened; yes is the default gesture. A rejection cannot come from surrender, because you needed an independent view to produce it. It is the approval gate's version of a certification eval: proof the check is able to fail. In a week of gate decisions, the rejections are the only entries that say anything about the operator.
Re-derive one approval a week
Pick one action you approved and rebuild the case for it from scratch, as if the recommendation never existed. The inputs, the alternatives, what you would have chosen on your own. Most weeks the agent's call survives the audit, and that is worth knowing. The week it does not survive is the week you find out what your acceptance streak was hiding.
So the rule I would hand any new agent operator: track one number, how many recommendations you rejected. That number measures you, and whether you still exist in the loop. If it sits at zero long enough, something has changed, and it is rarely the agent.
Some operational details in these essays have been changed for narrative or privacy reasons. The arguments, the numbers, and the lessons are real.