Sure, that’s an interface that’s better for many users and use-cases.

However, it seems better if you could, as much as is possible, move the AI stuff from runtime to “compile time.”

Instead of having the AI do everything all the time, have AI configure your Zapier (or whatever) on your behalf. That way you can (ideally) get the best of both worlds: the reliability and predictability of classical software, combined with the fuzzy interface of LLMs.

> Instead of having the AI do everything all the time, have AI configure your Zapier (or whatever) on your behalf.

That is what many use OpenClaw for! The AI assistant will happily recommend existing services and help you (or itself, if you let it), set it up.

(In theory. In practice, it often does a poor job).

The appeal of OpenClaw is I don't need to go research all these possible solutions for different problems. I just tell it my problem and it figures it out.

Yesterday I told it to monitor a page which lists classes offered, and have it ping me if any class with a begin date in March/April is listed. This is easily scriptable by me, but I don't want to spend time writing that script. And modifying it for each site I want to be notified for. I merely spoke (voice, not text) to the agent and it will check each day.

(Again, it's not that reliable. I'm under no illusion it will inform me - but this is the appeal).

That's still too much work. Someone would have to make like an OpenClaw wizard that protectively offers to set all that stuff up. So the potential OpenClaw user can then, on running for the first time, be guided through the setup of whatever they'd like to get connected. And "setup" here means a short description of X and a "Connect? (y/n)" prompt. Anything more and you start losing people.

yes. in a similar vein, we're seeing that get standardized in coding agents as "don't have the agent use tools directly, have the agent write code to call the tools"