I moved to Hermes after OpenClaw 2.0 failed to install.

That was the moment.

Not another review. Not a feature comparison. Not a long argument about architecture. I tried to install the release that was supposed to be OpenClaw’s great comeback, and it failed.

I had held on longer than I probably should have.

OpenClaw had given me something I still believe in: a personal AI agent that could live on my own machine, remember what mattered, work through Telegram, and remain open enough for me to understand and change it.

That promise was bigger than convenience. It was about privacy and freedom.

I held a torch for it.

The agents were getting better

The frustrating part was that the agents themselves had been making amazing progress.

Claude had become remarkably capable at taking a broad instruction and working through the details. Codex had become a serious builder. And recently, the Grok-powered bots had been especially impressive: fast, direct, willing to use tools, and much better at getting real work finished.

The future felt close.

I could see it in daily use. The models were becoming less like chat windows and more like colleagues who could research, write, inspect files, run commands, and carry a job across several steps.

That progress made the OpenClaw problems harder to accept, not easier.

The brains were improving quickly. The runtime around them kept becoming the thing I had to manage.

An agent is supposed to remove work. When updates, gateways, channel delivery, and configuration repeatedly demand attention, the assistant becomes another system on the maintenance list.

I had reached that point with OpenClaw before 2.0. I still wanted 2.0 to prove me wrong.

The release that was supposed to fix it

OpenClaw 2.0 was presented as the largest release in the project’s history. It had contributions from 933 people and more than 16,000 pull requests. The browser experience was rebuilt. Conversations could branch. Subagents were easier to watch. Shared sessions and stronger credential handling moved the product forward.

There is real work in that release. I am not dismissing it.

But I did not need a more ambitious OpenClaw. I needed a dependable one.

My agents lived in Telegram. I used them while moving around, between meetings, and away from the computer. I did not need to watch agents work in a beautiful browser interface. I needed to send a message and receive the result.

Most of all, I needed upgrades to stop feeling dangerous.

Then 2.0 failed to install.

That failure did not erase everything good about OpenClaw. It simply ended the argument I had been having with myself.

I moved to Hermes.

One request to Claude

The migration was almost embarrassingly easy.

I gave Claude one instruction: convert my OpenClaw setup to Hermes.

One instruction from me—not one hidden API call. Claude handled the actual steps.

I expected a project. I expected lost memories, broken personalities, missing skills, and a weekend spent rebuilding all the small decisions that had accumulated over time.

Instead, the important parts came across. The agents kept their roles. Their instructions had a home. Memories and skills had a path forward. Telegram came back.

It worked.

Perfect—not in the sense that software is ever perfect, but in the practical sense that matters. I asked for the move, Claude carried it out, and I could use my agents again.

For months, I had been asking Claude to fix OpenClaw.

Then I asked Claude to replace it.

The second request worked better.

Same models. At least three times faster.

The next surprise was speed.

Using the same models, my Telegram agents felt at least three times faster on Hermes.

That is my practical estimate, not a laboratory benchmark. I did not run a hundred matched prompts and build a latency chart. I lived with both systems, sent the same kind of requests, and felt the difference immediately.

Replies started sooner. Tool work felt immediate. The conversation stopped feeling like a request sent into another room.

Speed changes how you use an agent. Slow software teaches you to save up requests. Fast software invites you to think out loud, correct the work early, and keep moving.

The models had not changed. The system around them had.

The OpenClaw spirit

Charles Stross wrote in Accelerando:

“Software is not immortal.”

The software is not the spirit.

The OpenClaw spirit is the belief that an AI agent should belong to the person using it. It should remember. It should have a personality. It should live where you already talk. It should use your tools and work on your machines. You should be free to open it, change it, and move it.

That is what will survive.

I see it in Hermes. I see it in Claude, Codex, and the recent leap in Grok-powered agents. I see it in local models, open skills, and the refusal to accept that the most personal software we will ever use must belong to somebody else.

For KUON, Hermes is now the better home for that idea.

Something else may be better one day. If it is, I will move again. The agent matters more than the runtime, and the work matters more than the brand.

I did not leave OpenClaw angry.

I left after 2.0 failed to install.

Hermes gave my agents a better home. But the part of OpenClaw I cared about most came with me.

The OpenClaw runtime lost me.

The OpenClaw spirit did not.