2026.11.10
How to maintain critical thinking in the age of AI agents: sharing an article by Addy Osmani
In an age when AI agents do our work for us, what matters is not "doing the work yourself" but "being able to understand, judge, and take responsibility for what AI produces." Even if you leave generation to AI, humans decide what to build, what counts as correct, and whether it's finally OK to ship.
I learned a lot from Addy Osmani’s article on “how to maintain critical thinking in the age of AI agents,” so I’m sharing it!
In an age when AI agents do our work for us, what matters is not “doing the work yourself” but “being able to understand, judge, and take responsibility for what AI produces.” The more accurate AI becomes, the easier it is to accept plausible answers without question, and before you know it, you’ve let go of thinking itself.
That’s exactly why, before handing things to AI, you decide on hypotheses and completion criteria, and after execution, you check not only the test results but also the evidence behind them and the decisions the AI made. Even if you leave generation to AI, humans decide what to build, what counts as correct, and whether it’s finally OK to ship. That’s the gist.
More in the replies!
https://x.com/suna_gaku/status/2117966072120836302
① Delegating work to AI is different from letting go of thinking The article organizes how we engage with AI into stages: “Cognitive Offloading (delegating the work while staying responsible for the result),” “Deferral (taking the AI’s answer as the next premise without verifying it),” and “Cognitive Surrender (accepting the AI’s answer as your own thinking).” The danger is getting used to the convenience, gradually skipping verification, and at some point no longer being able to explain things yourself.
② Have your own hypothesis before asking AI If you ask AI for the answer from the start, once you’ve seen a polished answer, it becomes hard to consider other possibilities. It doesn’t have to be right, so write down your own prediction first. Then think about “what would prove this hypothesis wrong,” and check it against actual evidence.
③ Even if the tests are green, correctness isn’t necessarily proven When AI writes both the implementation and the tests, the same misunderstanding can creep into both, and tests can pass even for a broken feature. What matters is separating “what you were able to verify” from “what you weren’t able to verify.” Don’t treat passing tests as quality assurance in itself.
④ Look at evidence of what was actually run rather than the AI’s completion report Give the AI means of verification such as E2E tests, screenshots, logs, and traces, and have it run a fix → verify → re-fix loop. Before reading the completion report, check the commands that were run and their results. Also, if you define the completion criteria in advance, it’s easier to avoid conveniently judging something a success after the fact.
⑤ Review not just the code but the things the AI decided on its own Over a long task, AI makes many small decisions, such as default values, what’s out of scope, and how to interpret contradictory requirements. Since these aren’t visible from the diff or tests alone, have the AI list “the decisions it made and the assumptions behind them.” If needed, pick part of the code and check whether it can be explained why it was implemented that way.
⑥ The quality of a prompt reflects how deeply the human understands the problem Even with a vague request, AI can fill in the gaps and produce something plausible. That’s exactly why, before making a request, you should sort out the target users, what’s out of scope, the constraints, and the completion criteria in your own words. If repeated correction instructions don’t improve things, rewrite it starting from the problem definition rather than tacking more onto the prompt.
⑦ Don’t end AI failures with an on-the-spot fix; capture them in the system Reflect the reasons it didn’t meet the quality bar in shared settings, Hooks, tests, and so on. Rather than a human pointing out the same problem every time, make it so the AI can avoid it from next time on. By turning human judgment criteria into a system that can be reused continuously, you build up quality.
⑧ Use the amount humans can take responsibility for as the yardstick, not how much AI produces Even if you run multiple AI agents in parallel, human review capacity doesn’t grow at the same rate. Once the output exceeds what can be reviewed, strengthen verification or slow down generation rather than adding more agents. Also, once a week, use important skills without AI to keep your own judgment sharp. Much of the work and verification can be left to AI, but final responsibility can’t be delegated.