Agents produce fluent, confident answers at machine speed — exactly the signature Daniel Kahneman spent a career teaching us to distrust. If you ship with agents today, his fifty-year-old findings are your architecture doc.
In Thinking, Fast and Slow, Kahneman describes two modes of mind: a fast one that generates impressions effortlessly and a slow one that checks them reluctantly. The trouble is that fluency masquerades as accuracy — an answer that arrives easily feels true, and confidence tracks the coherence of the story, not the quality of the evidence. Anchors set early drag every later estimate toward them. Plans get built from the inside view, ignoring how similar projects actually went. None of this is stupidity; it is the default operating mode of any judgment system that optimizes for speed.
An AI-native studio runs System 1 in production: agents draft, code, and publish with total fluency and zero felt doubt. Kahneman's move — trust the process only where the environment gives fast, valid feedback — becomes a design rule. Put the slow system in the pipeline, not in your mood: evals that check shape before content, judges tuned for precision, base rates from past runs instead of the plan's inside view. Watch anchoring especially; the first draft an agent produces anchors every human review after it. Judgment isn't a vibe you apply at the end. It's infrastructure you build first.
- fluency is not accuracy, so never let an agent's confidence stand in for evidence
- make System 2 a pipeline stage — deterministic gates and judges, not end-of-day vibes
- estimate with the outside view, because your last ten shipping runs predict the next one better than the plan does.
