Fluency is not accuracy
Kahneman's fast system produces answers whose confidence tracks how easily they came, not how true they are. Agents work the same way: output arrives polished, plausible, and certain, whether or not it is right. The danger is not that agents err — it is that their errors read exactly like their successes. Reliable agentic work therefore cannot rest on how convincing the output feels; it needs clear task definitions, tests, and feedback loops that check results against reality rather than against tone. Treat every fluent answer as a claim awaiting evidence, and build the harness that demands the evidence, because no reviewer's intuition will flag confident prose as wrong.
Audit by stakes, not by vibes
The slow system's defining trait is laziness — it engages only when something forces it to. A founder cannot review everything agents produce, and pretending otherwise just guarantees the review is theater. The honest response is to ration scrutiny deliberately: match the approval layer to the stakes, so a typo fix and a wire transfer never share the same number of humans in the loop. Let agents own the wide base of reversible, well-specified, measurable work, and reserve human attention for the narrow tip — anything irreversible, high-trust, or hard to defend in a room with a customer. That is System 2 by design instead of by mood.
The planning fallacy now ships code
Kahneman documented how plans anchor on the best case: people estimate from the inside view, imagining the smooth path, and the first number spoken drags every later estimate toward it. Agents make this worse, because they generate plans and estimates with the same unearned confidence they bring to everything else, and cheap execution tempts founders to skip the outside view entirely. The counter is structural, not motivational — evals and tests that measure what actually happened, feedback loops that correct the next attempt, and a standing suspicion of any schedule that assumes nothing surprises you. Optimism is a fine fuel and a terrible instrument.
Design the pause before you need it
The deepest lesson of the book is that debiasing in the moment barely works; by the time the fast answer arrives, it has already won. The fix is to decide in advance where deliberation is mandatory. An AI-native company builds this in from the beginning — agents, workflows, evals, and human judgment designed together, not bolted on after the first confident mistake. In regulated domains like health, finance, and law, the human-in-the-loop is not a formality, and the escalation path must exist before you ship, not after. Humans keep goals, values, tradeoffs, and accountability. Kahneman would recognize the design: a fast system doing the work, and a slow one that shows up on purpose.
