BenGWeeks on Nostr: My experience is it could be do simple question / answer type stuff such as answering ...

My experience is it could be do simple question / answer type stuff such as answering e-mails, responding to messages, etc.

May be my expectations were a little high, but to do actual useful work it was challenging, e.g. creating presentations, etc. I really had to refine its skills and still then it struggled. It would forget what I had said to it, not finish tasks, etc.

I was using Opus 4.5 / Sonnet 4.5 (allowing it to choose models as appropriate). To reduce costs you can use a mix of models, reduce the heartbeat etc. Or use a local model. But I’ve found local models really struggle with tool use. That said, it was a costly experiment. Anthropic is just too expensive compared to what you get out of it (with OpenClaw).

That was a week ago though - and things move quickly atm.

Might have a go with ChatGPT (cheaper token use) and see if OpenClaw has improved.