Some agents don't call an API at all. They work the way a person does: open the app, click through the screens, type into the fields, press save. The screens are different, but the jobs are often the same ones a support or ops agent does.
A job that's clicked through the same screens every time repeats just like a job done through an API. That's why it's on our map.
Where we are
Clicking through desktop apps needs new pieces in AgentCompile, beyond the tool-calling jobs it compiles today. It's where we're heading. If your agent works this way, we'd like to hear about it.
The jobs are the same; the hands are different
A computer-use agent updates a record by clicking through an app instead of calling an API. The job is often one a support or ops agent would recognise: find the customer, open the account, change the field, save. What differs is how each step is carried out.
- Screens change when an app updates, so a step that worked yesterday may not work today.
- There's no clean answer to read back, only what's on the screen.
- Each click is slow and visible, which makes repeated jobs especially costly.
Why it's on our map
Those differences are why compiling desktop jobs needs new pieces in AgentCompile, beyond the tool-calling jobs it compiles today. The reason to build them is the same reason as everywhere else: the same job, done again and again, shouldn't be thought through from scratch each time.
Where we are, plainly
We don't compile desktop jobs today, and we haven't published any results for them. If your agent works this way, we'd like to hear what it does all day. It will shape what we build.