GM, quick tip:
if you are self hosting a private service like Immich,
make sure to give it a random subdomain, or a subdomain with a random suffix.
things like Immich might have exploits that havent been discovered yet, and if you give it a subdomain like `immich`, crawlers and bots might find it, and try to use that for exploits.
and in your DNS settings, use a wildcard, and resolve the subdomain privately, so nobody knows your valid subdomain list. and your subdomain is random. so even if there is a new exploit, nobody even knows the existence of your subdomain.
i think Cloudflare uses wildcard publicly for subdomains by default anyway.
few days ago i have been trying to make my own agent.
but a lot more isolated, permission based and also more free in thinking than clawd.
like you can see it type, it can decide to send the message or not, after typing it, like humans.
it can send multiple messages at once, or you can send multiple messages before it replies. it has a dashboard where it can see notifications and stuff. it can also update its own dashboard.
i just gave it an isolated filesystem, and ability to run typescript code with deno with limited file access permissions, can also spawn subagents etc. basic stuff, also some copy/paste tool, so it doesnt have to generate same text over and over again, can copy paste stuff in its context.
i didn't give it any web tools, or stuff.
gave it my social media, and told it to find as much stuff about me as possible.
- it created scripts it can use to crawl the web
- it created custom scripts for crawling some sites
- kept reporting me in the chat, and its feelings
- kept updating a nice md file with everything it learned about me, listed where it hit a wall, and what it can do about it, etc.
it was kinda cool.
my goal was making it async,
and but also making it like a coding agent with its own workspace, and stuff.
i know for memory people mostly use vector databases, but i think they are very limited.
so i also made a background agents, constantly reading the context of the master agent, and decide if there is something worth noting, and notes them in obsidian format, also tags them. then all agents can read from it, or add comments on them or modify them, like any other file, they can use vector search, or normal file search. as they walk the files, they add links to files linking to other files, like the obsidian graph. so agents can walk it and find out or remember stuff. same agent also injects relevant stuff to master agent's context, like "this might be relevant", etc. then master agent can decide to read it or walk it, or make some subagent walk it and summarize it to him. etc.
im thinking i might add more side agents, with different purposes. different personalities. noting stuff or telling stuff to the master agent.
idk, it was a one day thing, didnt touch it after that, i was in the middle of refactoring it. but maybe i rewrite it one day.
why lighting wins at scaling is the state.
every channel has an isolated state.
this is not only true for bitcoin, this a programming reality. even in physics and how the universe works.
so future of bitcoin scaling will be about isolated states. that nobody else has to download, and verify other than people who care about that state.