The premise for an always-on agent was a single computer
A conversation that asks someone who has actually run a continuously running agent what the prerequisites are. The conclusion lands not on the model or the prompt but on two things. One is a single dedicated machine that will not be cut off: close the laptop you work on and it stops right there, so you keep a computer switched on in a corner of the office and put the long jobs on that one. Used, it can be had for around ₩800,000, and the operating system does not matter. The other is a handover document: instead of tossing out a one-line ask to build something, you have to write and polish a requirements document together and hand that over for the result to come out dense. The standard for the document is whether it is helpful enough that a new hire could finish 90% of the work from it alone. The reason for splitting into several is context, not permissions; give each one only what it needs to know and efficiency rises while consumption falls.
Key points
- Keep a computer switched on at all times, separate from the machine you work on
- The result comes out dense only when the request is polished into a document first
- The standard is a document written well enough for a new hire to do 90% from it
- The basis for splitting is context, not permissions
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.