A Plan Document Can’t Hold Your Product Vision
Ras Mic’s core argument against fire-and-forget agent loops is an analogy: hire a very smart developer, hand them a spec, and let them build the entire product without consulting you. They will fill every gap with assumptions about look, feel, and architecture, and many will diverge from your vision. “You might think your plan document covers everything. But truth to the matter, it never does. There’s always an edge case. There’s always something that’s missed.”
An autonomous loop has the same failure mode, because a spec.md or prd.md is asked to carry judgment that hasn’t been formed yet. Anyone who has run a service business knows clients can never fully externalize what they want even to another human; an agent reading a document gets less, not more. The exception is work where the details don’t matter: Mic built an Among Us simulator to benchmark which AI models lie best in about ninety minutes with a single fire-and-forget instruction, and it got many details wrong that he never cared about. For meaningful products his conclusion is blunt: “Human in the loop is the best loop,” because “AI can replicate sauce. It can’t create sauce.”
Related
- Everyone Says Loop, Nobody Defines It — this is the substance behind the slogan fight: the loop people copy assumes a perfect spec nobody can actually write
- Three Modes of Working With AI — the same matching logic: fire-and-forget fits prototypes where details don’t matter, pairing fits work where they do
- Your automation isn’t set and forget, every agent needs a human on both ends — the same conclusion from the agency side: the framing human cannot be compressed into a document
- Agent Loops Only Work When the Output Is Binary — essay candidate: the missing fixed signal is exactly the judgment a spec can’t carry
- Stop Copying AI Workflows From People With Unlimited Tokens — essay candidate: the companion failure mode, cost instead of judgment