The Refrigerator Problem

At some point you need to stop improving the harness and start building the thing the harness is supposed to build. I know this. You know this. We all know this. And yet here I am, using Claude Code to set up hermes-agent, which is instructing NanoClaw to bootstrap itself, while I should be shipping features for shotclubhouse.com.

The dopamine hit from optimising this setup is so real and it is a proper rabbit hole. Every tweak feels productive. Every config improvement feels like it's going to save hours later. And it will. Probably. But at some point you're polishing the machine instead of producing value with it.

Anyway.

What Actually Happened

OpenClaw had to go. After Anthropic pulled the plug on third-party Claude access, I'd been running it on GPT-5.4 and MiniMax 2.7. It works but it lags. The coordination, the tool use reliability, the inbox processing. Everything is just a bit slower and a bit flakier than it was with Opus driving. The glitches I wrote about last week are real and they compound across a full day of work.

So the plan was: migrate to hermes-agent (the more sophisticated harness) and NanoClaw (which still works with Claude Max) running in parallel. Two harnesses, best of both worlds.

Geek Corner

Why NanoClaw still works with Claude Max: Anthropic blocks third-party harnesses by inspecting the payload. System prompt keywords ("claw", "hermes"), recognisable tool signatures (session_spawn, session_list), that sort of thing. NanoClaw strips the system prompt clean and renames tool declarations to something generic. Under the skin it's just the Claude SDK, same as Claude Code itself. Anthropic can't differentiate between a NanoClaw session and a Claude Code session because there's nothing in the wire format that gives it away. Could they put more effort into detection? Probably. But right now, stripped prompt + generic tool names = works like a charm on a Max subscription.

The Inception Chain

Claude Code sets up hermes-agent. As hermes processes each migration step (porting features from OpenClaw, setting up Fleet Lambda hardening, configuring Convex), it sends the equivalent setup instructions to NanoClaw. NanoClaw bootstraps itself based on what hermes tells it to do.

Three layers deep. Claude Code orchestrating hermes, hermes nudging NanoClaw, NanoClaw configuring itself. If NanoClaw's gateway fails, hermes restarts it. If hermes fails, Claude Code sorts it out. If Claude Code fails... well, I close the laptop and start thinking about life choices.

It's actually working. The GitHub Epics are flowing through. hermes and NanoClaw sides are done for Epic #1 (Convex). NanoClaw is working on Epic #1's remaining items (sleep cycles, discovery gossip, ACP metrics). The chain holds.

The Benchmarks Nobody's Written Yet

I'm running what is basically a controlled experiment across three setups:

1. OpenClaw (GPT-5.4 + MiniMax 2.7)
2. hermes-agent (GPT-5.4 + MiniMax 2.7)
3. NanoClaw (Claude Opus, vanilla, no custom intelligence)

Same tasks. Simple coding, complex agentic engineering, agent-to-agent coordination. The gap between NanoClaw vanilla on Opus and the other two is... significant. Opus on pure agentic work is just built different. The harness adds the final 20% but the model does 80%.

Then there's the second comparison: add wololo's agent intelligence layer on top. OpenClaw's new memory/losslessclaw/knowledge system. hermes's full self-improvement package. Custom intelligence across the board. Even with all of that, it still struggles to match NanoClaw vanilla on Opus for complex multi-step coordination.

That deserves its own long-form piece. The benchmarks are interesting and the conclusion is not what the "model doesn't matter, harness is everything" crowd expects.

The Point

Build the Coca-Cola, not the refrigerator. I'm going to keep telling myself this while I tweak one more config file.