Automating software through Citrix is computer use on hard mode

Most computer-use demos share an unstated assumption: the agent can inspect what it's operating. A DOM, an accessibility tree, at worst a local window handle. Citrix deletes that assumption. The application runs on a server somewhere else; what your agent gets is a video stream of a desktop and a pipe for keystrokes. Every element you 'click' is a guess about pixels.

This changes the engineering completely. You stop asking 'what is on screen' and start asking 'what do I believe is on screen, and how do I verify it cheaply before acting.' Every action becomes propose → act → verify, where verify means reading pixels again and being honest when they don't match your model. Latency makes it worse: the frame you verified against may be 400ms stale by the time your click lands.

What worked: treating the session like an unreliable network peer instead of a UI. Idempotent action sequences, checkpoints the agent can re-anchor to from any state, and a hard rule that writes get confirmed by reading back the result, not by trusting the click. The same discipline distributed-systems people learned decades ago, applied one abstraction lower than anyone wants to work.

The uncomfortable conclusion: if your agent can operate software through Citrix, everything with an actual API feels like a gift. I now read API docs the way other people read menus, with gratitude.