Wiki › Thinking next to AI

Critical Thinking with Agents

Practical ownership in agentic workflows: independent hypotheses, falsifiable verification, explicit decisions, review-capacity limits and recoverable understanding.

Critical thinking with agents means deliberately retaining an independent hypothesis, standards of evidence, understanding of consequential decisions, and accountability while delegating execution. In “Critical thinking when agents do the work,” Addy Osmani argues for agentic workflows, not retreat from them: as outputs become faster, smoother and less obviously wrong, the work itself stops forcing the human to think critically. More parallel agents still leave only one pair of human eyes.

Offloading, deferral and surrender

Osmani describes a gradual spectrum rather than a clean division between “using AI” and “not using AI”:

| Mode | What changes | Ownership |

| --- | --- | --- |

| Cognitive offloading | The AI does work the person intentionally delegates. | The person owns the result and its reasoning. |

| Deferral | The output becomes the next input rather than a hypothesis to investigate. | The person's independent frame weakens. |

| Cognitive surrender | The generated answer quietly becomes the person's own position without felt need to check it. | The person may no longer explain or defend the result, even with the agent's help. |

Full delegation can be appropriate when the cost of being wrong is low. The relevant boundary is consequence and answerability, not the percentage of code written manually. This also qualifies the comprehension-debt concern: a complete mental model of a large codebase is unrealistic; what matters is knowing which understanding is missing and being able to rebuild the relevant slice when needed.

Questions that preserve an independent view

• Before asking: write down a hypothesis, even a weak one, before a polished answer anchors the discussion.
• Choose the mode: am I collaborating to think, or intentionally handing thinking off?
• Seek a falsifier: what would have to be true for this conclusion to be wrong, and what actual test could establish it? A model's confidence is not that test.
• Hear a counterargument: ask for opposition from a different angle, then make one's own judgment rather than treating another generated answer as a verdict.
• Check answerability: would I stand behind this in front of the person who has to maintain it?
• Audit practice: which skills am I no longer exercising, and is that loss intentional?

Verification can itself be surrendered

Agent-run end-to-end tests, property tests and app interactions belong in Harness Engineering, but “green” proves only what a particular check actually covered. A test written with the same misconception as the implementation can pass while the feature remains broken. Osmani's review rule is to ask both what a check established and what it missed; if that boundary cannot be explained, obtain a substantive review rather than a quick approval.

Define “verified” before execution, then inspect the command and its actual output before reading the agent's completion summary. Prefer tangible evidence—tests, screenshots, logs or traces—over a declaration that the work looks finished. For research and product work, sample original customer calls or source documents and recalculate consequential numbers. These practices complement Production LLM Reliability; they do not turn one passing test into a general proof of readiness.

Review decisions, not only diffs

Generated summaries tend to hide small decisions accumulated during a session: defaults, scope exclusions and reinterpretations of conflicting requirements. Ask for those decisions and assumptions explicitly. Explain a few sampled lines, reconstruct a document's argument from memory, or use a quiz to reveal whether the result was actually understood. Review assistance can direct scarce attention; it should not silently set the human's standards.

Osmani calls the gap between agent-generated volume and work one can comfortably put one's name on the orchestration tax. If completed work grows faster than review capacity, another session is not the remedy: improve checks or reduce throughput. This is a human-capacity constraint on AI-Assisted Software Development, not merely a model-speed constraint.

Intent and standards belong at the boundaries

Write a task brief in one's own words: audience, constraints, exclusions and definition of done. An agent can fill gaps convincingly without resolving an unclear goal; Osmani calls the resulting hidden goals and rationale intent debt. After repeated unsuccessful corrections, rewrite the brief rather than endlessly tuning local instructions. An interview by the agent can help clarify intent, but cannot supply whose purpose the work should serve.

Attention shifts toward the front of a task—intent and context—and its end—evidence and review. Avoid chains in which each agent sees only the previous summary and nobody returns to the original. Encode reasons for rejecting output in shared instructions, hooks or tests so taste becomes reusable rather than merely cosmetic polish.

Deliberate calibration

Osmani suggests occasional unaided practice of skills one wants to retain, a monthly inventory of deliberately abandoned tasks, and treating verification time as real scheduled work. His weekly/monthly cadence is practitioner advice, not an experimentally established training prescription. Improved models can take over more detection and auditing; they do not decide which goals, trade-offs or consequences a human is willing to own.

This page records a single practitioner's normative framework, not independent validation of every anecdote, attribution or referenced research claim. Its practical synthesis is: delegate execution proportionally to risk, retain the ability to rebuild understanding, and separate generated confidence from evidence.