Every AI Agent Needs a Witness

Software that takes a task and finishes it on your behalf does not need to be smarter than a chatbot. It needs a witness: something, or someone, on the hook for whether the work it did is actually correct. That is the layer the AI industry has not built, and the gap shows up plainly in the adoption numbers.

The actual gap is institutional. Chatbots answer a question and stop; the user is the verifier. Agents act, and the user has no frictionless way to know whether the booking, the code, the report is right. A junior employee can spend an afternoon reviewing a draft. An automated tax filing does not get that review.

The Inlevel9 letter puts a number on it: citing OpenAI's 10M Codex/Work users against 900M ChatGPT weekly active users, the author calculates a 1.1% ratio — roughly one in a hundred chatbot users has touched an agent product. The same pattern shows up in the group-chat benchmark: if a product is not in the group chat, nobody trusts it enough to recommend it.

The reusable mechanism: every agent product will stall at the same wall until the question "who checks this?" gets a concrete answer before the question "how smart is it?" gets asked. The labs selling agents are optimizing the wrong variable. The winners will be the tools, human or model, that own the verification step.