Addy Osmani wrote a fascinating article titled " Critical thinking when agents d
By Brett Bohannon · November 5, 2026 · Curated by George's Blog
Addy Osmani wrote a fascinating article titled " Critical thinking when agents do the work" and the below is a summary of it and how he thinks about.
In the past, critical thinking happened in the space between question and answer. That space has collapsed to seconds. You still only have one pair of eyes, and nothing in the workflow forces you to think carefully. Pleasant, often-correct results make it easy to relax vigilance and slide into “cognitive surrender” (a term from Steven Shaw and Gideon Nave): the AI’s output becomes yours without real ownership.
He describes a spectrum:
Cognitive offloading (healthy): You delegate but keep ownership of the result.
Deferral: The tool’s output becomes your new starting point rather than a hypothesis.
Cognitive surrender: You no longer feel the need to check.
This creates “cognitive debt” or “comprehension debt” that stays invisible until you need to change the system (or explain a decision) and realize nobody (including you) actually understands it.
Couple Practical habits he recommends ( there are more)
Have a hypothesis first. Write down (even just mentally) what you expect before running an agent. It’s fine if you’re wrong; the point is to avoid being shaped entirely by a polished answer.
Ask yourself: Am I using the agent to help me think, or to offload thinking? Both are fine if you’re intentional.
Stress-test: What would have to be true for this to be wrong? Get counter-arguments from other models. Would you bet on this in front of the person who has to maintain it?
Treat verification as a habit. Green checks only prove one property. Define “verified” up front (tests, screenshots, logs, traces). Review what the agent actually ran and returned before reading the summary. Ask agents to list the decisions and assumptions they made.
For long sessions, have them quiz you.
Read diffs as if a human wrote them. Look past the summary at the small decisions (defaults, scope, reinterpreted requirements).
Watch the “orchestration tax”: if finished work piles up faster than you can stand behind it, you’re no longer reviewing—you’re hoping. Tighten checks or slow down.
You can (and should) offload a lot of work and even much of the auditing. You cannot offload accountability. Someone is still responsible for what ships, and the judgment of how much to trust an agent on any given task is itself a human skill that atrophies if you stop practicing it. The models will keep improving at detection, but questions of what to build, why it matters, who it affects, and what you’re willing to sacrifice remain yours.