But what I didn't understand... what benefit does it actually bring? On a default loadout, even after disabling tons of skills in the setup wizard, there were lots of useless garbage skills enabled, over 10k of context used just to list them all.

I vibe-coded my own harness that uses ACP so it supports any coding CLI that exposes ACP (copilot cli, opencode, basically every popular one with official or non-official wrappers). And I was able to achieve basically exactly what I wanted from any Claw-like agent, in like a few hours.

I know there's way, way more to these self-sufficient agents (compaction, memory system, etc) but in my mind, it feels like the closer you can be to a barebones "coding agent core" plus "gateways that point to it" the better.

OpenClaw has come so far since its' original launch/craze. I recommend taking a second look if you haven't touched it in a while. If for no other reason than it's just a really fun playground with a LOT of areas to experiment in. Setup a myriad of agents with various models, skills, cron jobs, etc. The control surfaces have come a long way as well.

Hermes is good fun, running that as well but feels like they focused on polish vs features in order to capitalize on the primitive state that OpenClaw was in for its first months.

People that got attracted to the hype of openclaw but couldn't endure the fast pace of breaking changes while they figured out the problem space were well served by going to Hermes.

I do have a website on which I've been adding more and more stuff for personal use and for sharing with people, but when I want to develop it I do it with any agent I'm using right now (Codex, Claude Code, Pi, etc) and when I want to ask questions about it, it's usually on the public internet so any chat interface can query it. That leaves two things: asking questions about more private stuff, and possibly a "claw" that lives on your computer/on a small server is less of a pain to connect to private stuff than building a MCP and authenticate yourself through it ; and apps that themselves use models, which can be developed by the "code agent" and then I can plug whatever model I want on it.

https://news.ycombinator.com/item?id=48187581

https://hermes-agent.nousresearch.com/docs/reference/skills-...

The purpose of these systems is not to be a system, it's to be a platform/substrate for agency expansion. Whoever can best expose their capabilities (without having to many tattering detractors) is ultimately going to be the winner.

I used OpenClaw -> Hermes -> my own thing now.

I've got things like code review, email inbox/spam filtering, website monitoring for bad links/typos, HN/Bluesky notifications, favorite director/actors/author alerts, etc.

I mostly interact with them using Slack and Email.

Both of them broke on me and were maddening to try and debug. I'd get them working then 1-2 days later send a message and get no reply. I went round and round with OpenClaw (seriously, "openclaw doctor" is a sick joke) before finally nuking the VM.

I want the simplest of things: Claude Code + My Obsidian Vault. That's it. I didn't even add other MCPs or Skills, I tried using Nano/OpenClaw and they both just fell over after a couple days.

I'm probably best served by just vibecoding my own tool since these things try to be too many things to too many people and end up failing.

So I looked elsewhere and found Crush and Hermes. They're both very a e s t h e t i c, which I think can make using them fun, but ultimately if I had a nitpick, I would look back over at pi, and the grass looked greener. (And not to mention, both Hermes and Crush seem to have some drama/baggage.)

I'm back on pi, and happy with just a few packages I've downloaded for it.

It only appears that the hosting providers are and not the user spending thousands of dollars on these agents.

I built smth similar which lets you delegate from say local model to Claude or mimicode to pi. It does audited, workspace segregated mcp calls with only 3 tools ever in context. It can store skills, cross harness memory, inbuilt AI first task manager. Workflows that can run on triggers or schedules which take skills, a model, a prompt and a hand picked set of mcp tools (very granular).

I use this to let me know what’s going on, to save tokens in Claude by delegating to MiMo and deepseek, and local models in pi.

No one using these things has a use case for them - if they did they’d have built it themselves like I have. Because with ai… you can!

Oh and it can work via telegram (cos if it doesn’t is it even an ai harness token furnace!!??)