You sleep about eight hours a night. Add a commute, dinner, the gym, the hour you spend on email pretending to work — and you’re at your desk maybe eight hours of every twenty-four. Your tokens, meanwhile, refresh on a clock that doesn’t care.
Most engineers I talk to are running their agents like a 9-to-5 job. That’s the wrong default.
The math nobody runs
Pick any agentic workflow — p0, raw Claude Code, Cursor on autopilot. Whatever your daily token budget is, you’re probably consuming it across an eight-hour window, hitting a rate limit somewhere in the middle, then closing the laptop. The other sixteen hours, the budget refills and nothing draws against it.
That’s 60-70% of your capacity, every day, walked away from. Across a team of ten engineers, it’s the cost of two more.
The reason it happens is mundane. Scheduling a build for later used to mean writing a script, leaving a process open, or trusting yourself to come back at midnight. Nobody does any of that.
Schedule the build before you close the laptop
In v1.0.50 we shipped Schedule Builds for Later. Every session has a “Start later…” button. You pick a time — tonight, tomorrow morning, next Monday — and p0 kicks the build off without you. No open laptop, no terminal session to babysit, no cron config to write.
The companion piece is the Schedule retry card. Hit a rate limit mid-session and you get a one-click option to pick back up exactly where the agent left off, automatically, the moment your quota resets. No copy-pasting context into a new session. No losing the partial work.
Both of these exist for one reason: to make 24-hour agent utilization the default, not a thing you have to remember.
Pick the right model for the job — overnight
We also shipped a per-session model picker in v1.0.50. Every chat now has a model and provider dropdown right in the header.
That matters more at 2am than it does at 2pm. Different models hit different rate-limit clocks. Some get cheaper or faster off-peak. If a session is going to run while you’re asleep, pick the model that makes sense for the session, not the one you happen to have selected when you scheduled it. Claude on the orchestration-heavy session, Codex on the long refactor, your call.
What overnight actually buys you
Wake up. Coffee. Open the laptop. Three sessions ran while you slept. Two have PRs ready for review. One hit an edge case at 4am, the retry card auto-fired at 6:15 when the quota came back, and it’s already merged.
That’s the morning we’re trying to build toward. Not “AI helped me write a function faster.” Sleep, and the work doesn’t stop.
The agents don’t need rest. Stop giving them any.
— Mario