the studio suite is the demo, but the thing that decides whether people leave it running is the agent's permission boundary, not the image gen. full filesystem access plus run-commands plus unattended scheduled tasks means the first time a daily job writes to the wrong path or pushes a half-baked file at 3am with nobody watching, trust is gone for good. what's missing for me: a per-action approval model (or at minimum a dry-run plus diff before any write) on anything destructive or outbound, and a dead-man's notification when a scheduled run fails instead of silently skipping. self-hosting the data is the right call, that's the part people actually want, but the approval and failure-visibility story is what makes someone trust it with real work. written with ai
the studio suite is the demo, but the thing that decides whether people leave it running is the agent's permission boundary, not the image gen. full filesystem access plus run-commands plus unattended scheduled tasks means the first time a daily job writes to the wrong path or pushes a half-baked file at 3am with nobody watching, trust is gone for good. what's missing for me: a per-action approval model (or at minimum a dry-run plus diff before any write) on anything destructive or outbound, and a dead-man's notification when a scheduled run fails instead of silently skipping. self-hosting the data is the right call, that's the part people actually want, but the approval and failure-visibility story is what makes someone trust it with real work. written with ai