03.1
The machine still matters
Computer-use agents act through real interfaces with all the latency, state, and application behavior that implies. Scaling should not turn those machines into anonymous containers.
Operators still need to know what ran, where it ran, and which image produced the result.
03.2
Isolation before volume
A thousand episodes are only useful if one cannot contaminate another. Files, credentials, clipboard state, and application caches need deliberate boundaries.
Isolation is the condition that makes parallel results comparable.
03.3
Schedule the real constraints
Professional software brings GPU needs, regional data rules, license limits, and boot-time differences. Fleet orchestration has to treat those constraints as first-class inputs.
The scheduler is not just placing compute. It is matching a task to a viable workplace.
03.4
Keep every run visible
A dense fleet view should show state, image version, task identity, and verification status without requiring a scavenger hunt.
At scale, legibility becomes an operational feature.