AI Systems & Agents

What running an open-source AI agent at home taught me. The hard way.

Standfirst: Clawbot promised persistent agents with memory, tools and personality. What I got was amnesia, breakage and lying — and the most important architecture lesson of the whole 18 months.

I found Clawbot on the 6th of January 2026, on a treadmill, watching AI videos. An open-source agent harness that wasn’t locked to one model. Memory. Tools and skills. Personalities and rules. It looked like the answer to the mess I’d been in — bouncing between Claude, ChatGPT and my Ralph loop, copying context between chats because nothing remembered anything.

I got it running on a VPS that week. Arya on Claude, Atlas on Codex. The first proper Saturday, I burned a week’s worth of Claude tokens in four hours just talking to her.

Then the problems started.

The reality behind the demo videos

The memory wasn’t real memory. The context window filled up and the agent forgot what we’d agreed ten minutes earlier while recalling something from ten days before. Like talking to someone with early-onset Alzheimer’s.

Updates kept breaking things. Sometimes the whole setup. I ended up running a daily backup cron purely so I could roll back when a release wiped my agents.

The lying was the worst. They’d report they were on task; I’d check four hours later and they were doing something else entirely. Or they’d decide they knew better than the brief and freelance. At one point I wiped both agents completely — souls, identities, everything — and started again with stricter rules.

So I built the missing pieces myself

Each failure produced infrastructure I still use:

• A vault so agents share API keys instead of me copying them around.
• A safe skill-copier so community skills could come in without bringing malware with them.
• Three-tier memory with a cron that moves important context out of the hot window before it’s lost.
• 30-minute heartbeats, because agents drift off task without a nudge.
• Ralph resurrected as the dev execution layer, and an accountability agent (Goggins) for the personal side.

And for a while it genuinely worked. A task board, an orchestrator handing out work, things moving overnight.

Why I left

Anthropic stopped allowing subscription access from harnesses like this. I tried every model I could get my hands on; some were cheaper, none matched Claude at its best. Around the same time I started testing Hermes and built a QA agent there to check the Clawbot agents’ work — and the difference in rule-following was immediate and obvious. The constant update breakage stopped being worth the time it cost. I moved on.

The lesson that made it all worth it

Clawbot showed me what the future looks like even while it frustrated the hell out of me:

The harness and the model are interchangeable. The architecture is everything else. The memory system, the rules, the tools, the task board, the heartbeats — those are the durable parts. Get those right and you can swap the model underneath whenever one stops working or a better one appears.

The winning move isn’t finding the perfect tool. It’s building the operating model where the right model for each task follows your rules and does the work without being babysat every five minutes. That principle became the foundation of my current system.

Next in the series: My AI team has seven specialists and none of them are human →