01 Platform · Scheduled runs
Work that happens with nobody at the computer.
Each agent has its own environment in the cloud, with a browser, a shell and a file system of its own. That is what lets a job run while the office is closed and arrive finished rather than described.
- Opened a browser and read the company site
- Pulled the last three jobs off the record
- Wrote the brief as a document
- Staged the email. Did not send it.
02 Goals, not just tasks
Keep working until it is actually done.
A task runs once. A goal runs until somebody independent agrees the goal is met, which is a different and more useful thing, and a much easier one to promise dishonestly.
A separate check decides
Not the agent marking its own homework. A different pass confirms the goal before the run closes as met.
You set the ceiling
An attempt cap and a hard limit on how long it may run. Open ended is not an option, on purpose.
A timeout is recorded as a timeout
Work that ran out of room is kept and reported as unfinished. It never gets written up as a success, because a run that quietly reads as done is worse than one that failed loudly.
A dead run resolves itself
If a run dies, it gets marked with the reason rather than leaving a schedule reading as if it never fired. Every run ends at a status you can read.
03 When things go wrong upstream
It held up when the provider did not.
When the AI provider underneath this went down for most of a day, the scheduled work did not fail in a loop and it did not silently skip. It paused, held its slots, and picked itself back up when service returned.
We know that because it happened rather than because it was designed for. An overnight outage became a morning where nothing needed cleaning up, which is the only version of that claim worth making.
And none of it went anywhere.
Everything an unattended run produces lands in the morning brief, staged, waiting on you. The point was never that something is awake at three in the morning. It is that it does not have to be you.