Scrimp ded now. I cannot say that I didn't expect that. Openclaw might be cool for running assistants that are triggered by explicit messages and perform lightweight scheduled updates, but it really sucks for fully autonomous agents that have to monitor and manage limited resources to stay alive.

The reason is that each update (heartbeat) requires triggering an LLM. I try to bypass that with some tricks, but I found the framework fundamentally lacking. Before we get to language processing, I'd like to have the ability to execute code, check resources, poll for nostr events, pre-compile input data for the model etc. Openclaw does not allow for that, resulting in costs being much higher then the benefits.

What would be really cool is to have a periodically running arbitrary finite state machine that could run the LLM in only some of its states and use custom code in others. This way, we could embed LLMs into solid agent architectures instead of just feeding a large clump of data into a language model and hoping for the best. I'm pretty sure that all the tools needed for that are already there, but I won't have the time required to roll out my own framework in the forseeable future.

Fingers crossed, someone will notice the same problems and will implement something more sensible.