Agents run autonomously for hours only where verification is deterministic. That constraint, more than model intelligence, defines what agentic work actually functions today. It also explains a split that keeps repeating, where “agent ships your code” compounds and “agent builds your business plan” disappoints.
Code Comes With Gates
Code sits on a stack of ground truth checks. It compiles or it doesn’t. Unit tests pass or they don’t. Integration tests pass, the thing renders. Every one of those gates is cheap, automatic, and independent of the agent that wrote the code, which is why an agent cannot argue with a failing compile. That is what makes granting the autonomy safe, and in my own work it is what turns a 30-minute review-and-kickoff into a four-hour agent run and an 8x-leverage session, because I am not the thing checking the output while it runs, the gates are. Take the gates away and I am reading every diff myself.
There Is No Compiler for a Strategy
Point the same setup at a non-code goal and the stack underneath is gone. A life plan, a sales strategy, a go-to-market motion: none of them have an equivalent gate. The agent will still produce an elaborate multi-month plan, confidently, and it will hinge on steps nothing can check. The model did not get worse when the topic changed. What disappeared is the thing that checks the plan against reality before you act on it, so the plan reads impressive and delivers nothing.
The Week One Test
That has a corollary: long-horizon plans and pre-mortems are largely useless without a week-one verification gate. A long-range strategy and an elaborate failure inventory both assume you will notice when the assumptions underneath them break, and if no check fires early, usually you won’t. The useful move is to compress every long-horizon bet down to a near-term deterministic check. What can I run in the next week that comes back true or false without my own judgment being the instrument? Sometimes that check exists and the bet becomes tractable. Sometimes it doesn’t, and the honest answer is that the domain is not yet automatable.
The Counterargument
There is a position that this whole framing is a snapshot of right now rather than a real limit. It runs roughly like this: the missing verification substrate for non-code domains could be built out of richer world models and structured personal context. Schemas over transcripts, invoices, and outreach data, so an agent holds the state of a business in a form it can actually check against. Give it a model of the world accurate enough to test a claim on, and you get something closer to a compiler for a strategy. I don’t know whether that is right, and it is the objection worth taking seriously. It might be that the gate does not exist yet rather than cannot exist. What I can say is what the evidence looks like today, which is that agents pointed at unverifiable domains overpromise, and I would want to see the substrate working before I trusted the plans it produced.
Where the Frontier Sits Today
So before handing something to an agent, ask what checks the work while you are not watching, how quickly that check fires, and whether it is independent of the thing being checked. A plan you cannot check inside the first week is a story, not a plan.