Kahneman spent a career proving that fast, confident answers are not the same as correct ones — and every agent you ship is, by architecture, a fast-answer machine.
Thinking, Fast and Slow maps the two modes of human cognition not as a metaphor but as a predictive model of error. Daniel Kahneman shows that the fluent, automatic mode — System 1 — is not lazy thinking, it is the default, and it produces coherent stories out of incomplete information with no internal flag that anything is wrong. Overconfidence, anchoring to the first number in view, and systematically underestimating how long things take are not bugs in bad thinkers; they are the factory settings of judgment under uncertainty.
An LLM-powered agent is System 1 at industrial scale: fast, fluent, and constitutionally unable to feel uncertain. It will anchor on whatever framing appears in the prompt, construct a confident narrative from partial context, and hand you a delivery estimate that ignores every comparable project that ran long. The founder's job is not to trust the output less — it is to architect the System 2 that never shows up on its own. That means structured checkpoints, adversarial review steps, and explicit slowdown triggers baked into the workflow before the agent ever touches production decisions.
- Fluency is not accuracy — treat confident agent output as a hypothesis, not a verdict
- anchor effects travel through prompts, so audit the framing your agents receive, not just the answers they produce
- the planning fallacy will colonize any timeline an agent generates unless you force a reference-class comparison into the process itself.
