results of the first try, building an agent from scratch by prompting (the task = preparing linkedin applications for a few jobs):
the agents make costly mistakes. it took three attempts until it started handling "apply" button and use my knowledge library - despite i gave a clear detailed instruction about that from the very beginning. every rebuilding attempt took a few hundred creds, besides tests. i don't know if my original instructions were badly prompted or the production agent just ignored them (well, that's bad prompting too).
computer use is very cumbersome. running a test with their 500 free creds did not make it. it took ~150 credits first (kinda reasonably), then went in a loop, burnt ~400 more and suddenly stopped. and it was with claude sonnet, a clever model, so the cheaper ones should behave even worse.
computer use is embarassingly slow (a test run took at least an hour), and i didn't see how could i correct the agent's steps on the run anyway, so just had to let it go and eventually fail as is.
overall, this platform doens't look newbie-friendly at all. maybe for trained professionals it would become more usable and even cost-effective, let's see..
p.s. ah, and usual problem of LLMs - hallucinations - is flourishing here. i asked to use my knowledge base, it didn't, and started dreaming up instead, without asking
