Why I Left OpenClaw and Hermes for Something Way Simpler

I started with OpenClaw the week it came out and it instantly blew my mind. I remember thinking, we're going to talk about life before OpenClaw and life after OpenClaw. And I still believe that. AI was already awesome, but after going deep with OpenClaw I felt like I'd been running a race on one leg and somebody just handed me the other one.

I'm extremely grateful for what Peter Steinberger and OpenClaw did for me. It showed me what was possible. It opened the floodgates.

And it was great. Until it wasn't.

One day I looked up and realized OpenClaw had stopped helping me. I had turned into a developer maintaining a massive, bloated system.

And I'm not a developer. I know zero code.

The Moment I Knew

I spent all day debugging. It got extremely slow and sluggish. All the bloat piled up and everything took forever. Then every update meant 30 minutes to an hour of fixing things before I could get back to actually using it.

So I'd sit there fixing my assistant instead of my assistant fixing stuff for me. That's backwards. The whole point was to never touch this stuff again.

Then I Tried Hermes

I liked it. For a while. Then it started turning into the same problem. The "one agent for everything" idea just became one massive bloated mess again.

Hermes also had this "self improving" thing that I was really intrigued by. The agent tweaks its own skills and gets better on its own. Cool idea. But it turns out I'm too much of a control freak for that. I want to choose what goes into my agent's skills. I want to know exactly what it's going to do and why.

To be fair, I haven't used OpenClaw or Hermes in a while. Maybe they're both better now. This is just what happened to me.

What I Run Now: Pi

Pi is a minimal agent harness. Their whole pitch is one line: "Adapt Pi to your workflows, not the other way around."

That was the polar opposite of what I had. Ultra minimal, but you can customize it to do literally anything. I wanted to lean into that hard.

Fun fact I didn't know going in: OpenClaw was built on top of Pi. For almost five months, OpenClaw's engine literally was Pi's agent core. They swapped in their own engine in late May, but the README still thanks Pi's creator. So I didn't switch to some random new tool. I went one layer down to the engine OpenClaw started with, without the platform on top.

The part I love most is control. Pi's whole default prompt is about 600 tokens and it's open source, so I can read every line of it, and I can swap in my own with one text file. If an update ever touched it, I'd see the diff in seconds. Nobody is quietly turning my agent into something else one morning.

The Trick That Made It Click: Profiles

I went back and forth on this one. Keeping everything in one chat seemed simpler. One agent, one place, done. But that's exactly what turned into a bloated, confusing mess twice already.

So I split my agent Fred into profiles. One for SEO. One for code. One for content. One for finance. Each one is its own little specialist. Pi doesn't have a profiles feature, by the way. Mine are just folders and a launcher script.

Here's how each profile is set up. This is way less technical than it sounds.

Why this works: my SEO profile doesn't need to know anything about my finances. So it never loads that stuff. Less context, fewer skills, faster answers. Every lane is laser focused instead of one giant bot that's kinda okay at everything.

Each skill is as lean as I can make it too. A skill is just a text file that tells the agent how to do one specific job. When it's small, it doesn't eat up the context window, and the agent stays fast. It saves you on tokens too.

What's Actually In Mine

This is not some huge setup. It's ridiculously simple:

piece	count	what it does
extensions	3	a compaction tweak, a fast mode toggle, a hook for my terminal manager
skills	~20	coding, audits, security, site QA, research, brand + voice guides
prompt templates	3	audit, wrap up the session, "explain that like a human"
search tools	2	web search + page reader

That's it. Pretty bare bones. But if I ever wanted to tweak it, I can do literally anything. That's the part I love.

"Can't You Do All This With OpenClaw?"

I know what you're thinking. OpenClaw has skills. Hermes has skills. Both let you run more than one agent with its own memory. Hermes even calls them profiles. On paper, yeah, you can build most of this on either one.

The better question is what you have to carry around to get it.

OpenClaw and Hermes are platforms. They ship with messaging bridges, memory systems, agent routing, a hundred features you might use one day. All of that loads whether you use it or not. All of it can break on an update. And all of it was written by someone else, so when something acts weird, you're digging through their system to figure out why. I remember how annoyed I'd get having to restart the gateway over and over. I've never once had that problem with Pi.

Pi ships with almost nothing on purpose. No sub agents, no plan mode, no built in memory, no chat integrations. Their docs literally list the stuff they left out. Every feature in my setup is there because I put it there.

Here's what that looks like in numbers. I pulled these from the package registries myself.

	openclaw	pi
dependencies	65	19
package size	~184 MB	~22 MB
releases (last 3 months)	52	33
system prompt	built by openclaw, you add files	one text file, replace the whole thing
chat apps, cron, memory	built in	add it yourself if you want it

Here's the number that made it real for me. I installed all three fresh, changed nothing, pointed each one at a fake endpoint that logs what it sends, and typed "hi". Same message, same tokenizer. This is what each one loads before you've asked for anything:

stock install	tokens on "hi"	tools loaded
openclaw	32,682	51
hermes	11,334	20
pi	1,349	4

One note on the numbers: I counted with OpenAI's tokenizer, so your provider's exact count will land a little different. The gap won't. Three quarters of OpenClaw's number is tool definitions. Fifty one tools, every one of them described to the model on every single message, whether you're asking it to restart a server or saying good morning.

My tuned Pi lanes land between 5,000 and 12,000 depending on how many brains that lane carries, and I can still trim them. But stock Pi at 1,300 is the floor. That's how little the harness itself costs. Everything above it is stuff I chose to load, so I know exactly where every token comes from, and exactly where to cut.

I have never once had to debug a Pi update. Their changelog does list breaking changes, but the recent ones are all aimed at people writing extensions or plugging into the SDK. If you're running a few text files like me, the updates can't reach you.

"But I'm Not Technical"

Neither am I. Zero coding. It did take me a few weeks to get everything dialed in, but I also got a little obsessive about it. The actual process was simple, even for a non coder.

I just told the AI what I wanted. Or I told it what I was thinking and asked what it thought, and we brainstormed. Then it built it. No code from me at any point.

If OpenClaw Is Feeling Slow For You

Stuff was super buggy for me on OpenClaw and Hermes. I've had zero issues with Pi. If yours is feeling slow and janky, try this:

That's the bare bones version. No rules file, no skills, nothing. Stock Pi is honestly pretty awesome on its own. It's fast and it's smart. Then you can customize it to do and act however you want.

I'm still always experimenting. I still use Claude Code for some stuff, it's great at orchestrating and running sub agents. I mess around with Codex some, it's really good for computer use. But Pi is my main harness now, and I'm absolutely loving it.

<< BACK TO POSTS