Agentic tools have compressed the distance between a question and its answer to a few seconds. That interval used to be where most reasoning happened. This essay argues that the main risk of agentic work is not poor output but good output: fluent, mostly correct results that give the reader no reason to look twice. It proposes a three-stage model of delegation, examines how verification and review can themselves be surrendered, and closes with a small set of practices for retaining ownership of what one ships.

1. The problem of the pleasant answer

Critical thinking was traditionally taught as a disposition: meet new information with curiosity and measured doubt, weigh competing views, and then decide. That disposition was reinforced by friction. Finding an answer took time, and the time forced thought.

Agentic workflows remove the friction. As models improve, obvious errors become rarer, so the cues that once triggered scrutiny disappear. The work is also pleasant, which makes vigilance harder to sustain. And however many agents run in parallel, the output still passes through one reader.

The consequence is simple. Nothing in the workflow will impose critical thinking by default, so it has to be designed in.

This is not an argument for using agents less. It is an argument about where human attention should go once they are in use.

2. A spectrum of delegation

Three positions are worth distinguishing.

The third term, cognitive surrender, comes from Steven Shaw and Gideon Nave at Wharton. The movement down the spectrum is rarely noticed because each step is small and defensible: one careful reading earns a skim of the next answer, then a glance at the outline of the one after. The drift is discovered only when someone asks about a specific line and no explanation is available.

The accumulated cost has been described under several names. Margaret-Anne Storey writes of teams losing their shared theory of a system. In code, the equivalent is comprehension debt: a codebase that looks clean and passes its tests but that no one can safely change. The same failure occurs when a strategy document is approved on the strength of its table of contents.

Two qualifications matter. First, holding a complete mental model of a large system is unrealistic, and unnecessary. Understanding can be rebuilt locally and on demand, provided one knows which parts are missing. Second, full delegation is appropriate where errors are cheap. Choosing between three comparable hotels is a reasonable thing to hand over. Deciding what to build, what is safe to release, or how to invest is not. A workable rule: delegate where being wrong costs little, and keep the reasoning where other people absorb the consequences.

The practical test of ownership is explanation. If an answer cannot be explained, even with the agent's assistance, it has been surrendered.

3. Questions before and after the prompt

The older checklist (the right problem, the right method, the evidence, the stopping point) still applies. Agentic work calls for a revised version.

1. Was a hypothesis formed first? A prior guess, even a poor one, protects against the anchoring effect of a well-written response.
2. Is the agent helping the thinking or replacing it? Either is acceptable. Not knowing which is the problem.
3. What would make this wrong? A test that can be run is worth more than a model's stated confidence.
4. Has the opposing case been heard? Asking more than one model to argue against a firm position is slow, but instructive.
5. Would this be defended in front of the person who must maintain it? This is the standard of answerability.
6. Which skills are no longer being exercised? Letting some go is fine. Letting them go unnoticed is not.

4. Verification, and its counterfeit

The bottleneck in software has moved from writing code to establishing that it is correct. Agents equipped with a means of checking their own work (end-to-end tests, property tests, direct interaction with the running application) can iterate before a human looks at anything. Building those loops, and reviewing the evidence they produce, is now a core part of the job.

Verification can be surrendered too. "Is there evidence?" quietly becomes "is the check green?" Tests are cheap to generate, and a test written with the same misconception as the feature will pass. A passing check establishes one property and is easily mistaken for a general assurance.

Three habits follow:

• Define "verified" in advance. Criteria set after the fact tend to fit whatever happened.
• Require tangible completion evidence: a passing test, a screenshot, a log. Read the command and its output before the summary.
• Ask what the check did not prove. If that question has no answer, a full review is warranted rather than a quick approval.

The principle extends beyond code. A synthesis of fifty customer calls deserves a spot-check against a few source transcripts, and any number that carries a recommendation deserves to be recalculated once.

5. Review as attention

Consider a 600-line change. Its structure is sensible, names are clear, tests pass. Somewhere in the middle, a transaction boundary has been reordered or a default inverted. A review that looks only for expected problems will approve it.

Reading a diff is necessary but insufficient. Over a long session an agent makes many silent choices: defaults, scope, and which requirement gives way when two conflict. None of these appear in a summary reporting success. Asking the agent to list its decisions and assumptions makes them visible. Self-testing then confirms comprehension: pick lines at random and explain them, or close a document and reconstruct its argument from memory.

Volume is the remaining constraint. The gap between what agents produce and what one would willingly sign can be called an orchestration tax. When finished work accumulates faster than it can be examined, another session does not help. The options are stricter automated checks or a slower pace. Past that point, review has turned into hope.

6. Input quality and intent

Human judgment now sits at the two ends of the task: the brief and the review. A vague request to a colleague usually provokes a clarifying question. An agent tends instead to fill the gaps and return something that looks complete. The unstated goals, constraints, and reasons constitute intent debt, and unlike other debts, it can only be repaid by the person who holds the intent.

A prompt is therefore diagnostic. If the problem cannot be stated clearly, it is probably not yet understood. Useful measures include writing the brief in one's own words (audience, exclusions, definition of done), rewriting rather than patching after two failed corrections, and asking the agent to conduct an interview about what remains unclear. Andy Clark's distinction between delegating to a system and collaborating with it applies here: in collaboration, each exchange sharpens the person's model of the problem as well as the output.

Two cautions. In multi-agent chains, summaries of summaries replace the source; the third agent may never see the original document. And since every output now arrives polished and confident, polish has stopped being evidence of quality. Personal standards have to do that work, and recording why something failed them, in configuration or tests, is how those standards accumulate.

7. The objection from better models

If models will soon detect everything a human reviewer can, why maintain the skill?

Much detection will indeed shift to agents. Three things do not shift with it. Responsibility for what ships remains with a person. Deciding how far to trust an agent on a given task is itself a judgment, and it stays calibrated only through contact with the work, because the pattern of model errors keeps changing. Anthropic's research on its own staff names this the paradox of supervision: "supervising Claude requires the very coding skills that may atrophy from AI overuse." Finally, the questions of what to build, why, for whom, and at what cost are not detection problems at all.

The work and much of the auditing can be handed over. The accountability cannot.

8. Practice

None of this requires rigor at every step. A modest routine is enough:

• Weekly: do one piece of real work without an agent, as calibration.
• Monthly: list what has stopped being practiced and decide what may be let go.
• Per task: note a prediction beforehand and, afterwards, one thing the agent did not check.
• Always: treat checking as scheduled work with a cost, and learn what is not understood instead of pasting a one-line gloss.

Conclusion

Agentic tools change where thinking happens, not whether it is required. The interval between question and answer has nearly vanished, so the reasoning that once filled it must be placed deliberately: in a hypothesis before the prompt, in evidence before the summary, and in a standard applied before release. The governing test is plain. If the result cannot be explained and would not be signed, it is not ready to leave.

---

AI Engineering Fundamentals (Free to enroll)

Start building with language models: understand tokens and costs, design and test prompts, get validated JSON output, handle errors and safety risks, and finish with a guided document assistant project.

Machine Learning Fundamentals (Free to enroll)

Comprehensive Machine Learning Fundamentals (with Python) certification program covering core algorithms, data preparation, predictive modeling, model validation, performance optimization, and real-world machine learning applications.

Python Programming Fundamentals (Free to enroll)

Learn Python from zero in your browser, no installation needed, covering variables, decisions, loops, collections, functions, files, JSON and error handling, finishing with a real project that cleans a messy expense export into a trusted report.