Mollick's New Framework: The Question Isn't Whether AI Can Act Alone. It's Whether It Knows When to Stop.

Before you approve an AI agent for any task with real consequences, there's a question worth asking that has nothing to do with how capable the underlying model is: does it know when to stop and hand the decision to you?

Ethan Mollick's essay "Agency and Agents" (August 31, One Useful Thing) gives that question a name. He borrows "dark factory" from manufacturing — a plant that runs with the lights off because no person needs to be on the floor — and sets it against the fully manual alternative, where a person supervises every step. His argument is that almost none of the useful agent work happening right now sits at either extreme. It sits in what he calls the "twilight factory": a facilitator agent that keeps running the work but is built to proactively route a decision to a human the moment that decision would actually change the outcome. Not "check in periodically." Not "flag anything unusual." Route the specific decisions that matter, and only those.

That reframes the design work for anyone deciding whether to turn an agent loose on a task. Picking an autonomy level from a vendor's dropdown menu — "supervised," "semi-autonomous," "full" — isn't the decision that protects you. The decision that protects you is writing down, before the agent runs, the specific moments where a wrong call costs something real, and building the checkpoint there instead of somewhere more convenient or more generic. A checkpoint on everything is the same as manual work with extra steps. A checkpoint on nothing is a dark factory you didn't mean to build.

Picture a marketing operations manager at a 60-person B2B SaaS company who sets up an agent to monitor ad spend across four platforms and shift small amounts of budget toward whichever campaign is converting best that week. Most of that is safe to fully automate — a five-percent shift between two campaigns that were both approved for spend is exactly the kind of low-stakes, reversible call the twilight factory is built for. The checkpoint belongs somewhere specific: the moment the agent wants to pause a campaign entirely, or move budget toward a new audience segment nobody signed off on, because those are the decisions that change what the company is actually doing, not just how efficiently it's doing it.

Now picture a legal operations coordinator at a 25-attorney firm using an agent to triage incoming contract requests — routing routine NDAs to a template, flagging anything with unusual indemnification language for a lawyer. The failure mode here isn't the agent moving too fast. It's the agent being told to "flag anything unusual" without anyone defining what unusual means in a contract, which is exactly the generic checkpoint Mollick's framework argues against. A facilitator agent that escalates by vague instinct will escalate everything or nothing, depending on the week. One that escalates specific, named conditions — an indemnification cap above a stated threshold, a governing-law clause outside two approved states — actually does the routing work it was built for.

The twilight factory isn't a new capability you need to buy. It's a name for the design work you already owe any agent you deploy: decide exactly which calls are worth a human's attention, and build the agent to stop there — nowhere else.