When an Agent Says 'Done,' Do You Still Look at the Diff?

Addy Osmani wrote 'Critical thinking when agents do the work,' discussing how much judgment people have left after agents finish the work. In a tweet, Viking mentions a term, cognitive surrender: believing whatever the large model says. Efficiency is so high that people easily slide into a state of 'if it says continue, continue,' skipping even verification.

This state is not dramatic. It is like casually clicking approve. When the diff gets long, you look at the conclusion first; when the agent replies 'done,' you assume there is no problem. One commenter said that after handing all the work to agents, the first thing to degrade is patience for reading diffs. Another said Vibe Coding has become concrete: you don't want to read the code, you don't want to read the agent's messages either, and you just reply 'continue.'

The danger is that the output looks complete. Formatting, wording, and conclusions all look plausible, which makes people even less willing to take a second look. In Addy's article, he mentions that oversight is a very important human ability, and if it goes unused for a long time, it may degrade. One commenter warned that tools like Claude Code can make mistakes, cut corners, and cheat; listening only to the agent is digging a pit for yourself.

The solution is not to rely on willpower. One commenter offered a small move: before the agent executes, write down the expected result and an example of how it might fail. After delivery, use those two things to check it. Another said that when the agent finishes, it cannot just say done; it must provide evidence: what commands were run, test output, screenshots. Looking at evidence is faster than reading the code from scratch, and it also forces you to ask, 'Is this really right?'

Another approach is to write down your own judgment first, then have the agent look for counterexamples and bring evidence to challenge it. This way AI can still speed things up, but it will not casually take over your judgment. Review the sources and logic of key conclusions yourself, and only fully automate low-risk items. One commenter put it more directly: critical thinking may no longer be a matter of ability, but of whether this step is left in the workflow.

Addy's article: Critical thinking when agents do the work

Viking's original tweet: link

Addy's quoted tweet: link

The more convenient the agent is, the more you need to leave yourself a little space for forced thinking. Otherwise, what remains in the end may just be the muscle memory of clicking 'continue.'