Tokens Need a Critique Loop¶
Jensen Huang's line — a $500k engineer who only spends $5k on tokens is underutilizing coding agents — is directionally correct but insufficient, per mikhail-parakhin. Raw token consumption correlates with output only when paired with a structured critique loop: one agent writes, a second (ideally a different model) critiques, iterate until the second accepts. Latency goes up, which people resist. Quality goes up, which is the point.
"It's not about just consuming tokens… The anti-pattern is running multiple agents too many agents in parallel that don't communicate with each other. That's almost useless compared to just fewer agents and burns tokens very efficiently."
The parallel-agent anti-pattern is specifically important: spawning N agents that don't read each other's work consumes tokens without accumulating any signal. Two agents in a critic–author dialogue beats ten running in ignorant parallel.
Why it matters¶
Pairs directly with ai-generated-code-is-untrusted and do-not-outsource-thinking: the critique is the thinking. Without the critic, volume just dumps more bugs into prod — Parakhin's claim that models now write code with fewer bugs per line than a human, yet produce more total bugs in production, because they write so much more of it.