The short version: In one week, OpenAI and Anthropic both shipped agents whose headline feature is a unit of time. OpenAI's ChatGPT Work, launched July 9 alongside the general availability of GPT-5.6, is described as an agent that can "stay with a project for hours" and take action across your apps and files. Anthropic's Claude Cowork runs scheduled, asynchronous work while your laptop is closed. MiniMax pitches an agent team built for long-running tasks. None of these launches led with a benchmark. They led with duration.
That's the tell. The spec that moved this week wasn't a price or a leaderboard row. It was how long the thing runs without you.
Why hours is a different product, not a bigger one#
A single-turn assistant hands you text. You read it, you decide, you act. The model's mistakes are contained inside a reply you inspect before anything happens in the world.
An agent that works unattended for two hours inverts that. It acts first — dozens of times, in sequence, each step conditioned on the last — and you inspect afterward, if at all. Between "go" and "done," it may send emails, open tickets, edit records, move money. The intelligence didn't necessarily get better; the amount of consequence you delegated without looking got bigger.
The variable that grew this week wasn't the model's IQ. It was the length of the rope. And a longer rope changes what you have to worry about, not just how much you can get done.
This is the part the launch demos skip. A wrong turn in minute three of a two-hour run doesn't announce itself. It compounds — the agent builds its next twenty steps on the bad one — and you meet the result after the blast radius is already the size it's going to be.
The controls that matter are now operational#
Here's the uncomfortable reframe for founders: once you delegate hours, your reliability stops coming from model choice and starts coming from operational guardrails you build. The lab shipped the capability. The blast radius is yours to bound. Four controls, in rough priority:
- A hard spend cap. Not a dashboard alert — an enforced ceiling on tokens and dollars the agent cannot cross, because an agent looping unattended is exactly the thing that runs up a bill you find on the invoice. This is a solved problem; enforce a token budget before you hand off, not after.
- Approval gates on consequential actions. The agent can research, draft, and plan for two hours freely; the moment it wants to do something irreversible — send the email, charge the card, delete the record — it stops and asks. Cowork's own design leans this way: it leaves work drafted-but-unsent and pings you for the decision that's yours. Copy that.
- Durable recovery. A two-hour run will meet a crash, a redeploy, or a rate-limit wall. Without durability, that means lost in-flight work or — worse — repeated side effects on restart. This is the dual-write problem and the idempotency problem at the scale of a whole workday, and it's why where and how you run a long agent now matters more than which model drives it.
- A readable audit log. After an unattended run, "what did it do" has to be answerable in seconds, not reconstructed from token traces. Every consequential action, timestamped, in a form a human skims.
The founder move#
Don't spend this week benchmarking Sol against Sonnet against Gemini for the hundredth time. The models are close and getting closer; that race is not where your risk is.
The move is to ask the question the demos don't: what is the worst thing this agent can do in the two hours I'm not watching — and have I capped it? If you can't answer that, you haven't adopted long-horizon agents yet. You've just pointed a faster machine at your business and hoped. The labs made the hours cheap. Making the hours safe is the part they left on your desk — and it's the part that decides whether "an agent that works while you sleep" is leverage or a liability you discover at 9 a.m.
