Independent research / 2025-2026
Persistent agents at home.
Persistent agents, run on home lab for months, to find out where they actually earn their keep and add value to my life.
Not a demo and not a review. I gave the agents memory, schedules, and triggers, kept them running on an isolated Linux VM, and watched what they were useful for — and what they were not... and how and when they broke.
The premise
Persistent agents with memory, schedules, and triggers should be useful outside a demo. I ran a handful of agents in different tools for months to find out where they earn their keep (token spend and maintenance time).
My role
Everything: setup, operations, repair, migration, evaluation.
The result
Four use cases that survived, a stability lesson that moved recurring work into n8n, and two frozen agents on a dormant VM.
01 / The problem
Demos are easy. Keeping an agent running is not.
Persistent agents are the useful promise of the current wave: software with memory, schedules, and triggers that keeps working after you close the window. The demos make that look settled. I wanted to know what it takes to run one at home, for months, on real work.
So I treated it as fieldwork, not a review. The agents got their own context and stayed on a Linux VM. I paid attention to where they earned their keep and where they did not.
02 / What I built
Two agents, run like a product.
The first was OpenClaw, in February 2026, early in the craze. Claude helped with the setup. For a couple of months it ran custom skills and cron jobs, and I spent a lot of time experimenting with what my “OpenClaw buddies” could do. ChatGPT OAuth worked easily for both agents, which made a fun parallel track while I was learning Claude Code and then Codex; some scheduled agentic tasks eventually moved into ChatGPT Work.
The use case that earned its keep first was testing AI SkillsBank as a real user. The agents had their own browser and email. They signed up, worked through the features, and sent back feedback. They also ran research radars across enterprise and consumer AI and robotics, and wrote a weekly narrative on my sports teams from game reports and transfer news. The last job was an instructive one: moving recurring agent behaviors into n8n.
03 / What I learned
The agent is the easy part. The upkeep is the product.
OpenClaw needed constant repair. Migrating to Hermes Agent carried baggage from the old setup and did not feel better. Starting a fresh Hermes agent, with no baggage, went much better. Fresh context beats migrated baggage.
The deeper lesson was about stability. OpenClaw is always breaking and Hermes is always learning, so anything that has to stay stable does not belong in an agent. The recurring behaviors moved into n8n: more stable, cheaper, and boring in the way production software should be.
Anything that needs to stay stable belongs in a deterministic workflow, not an agent.
04 / Why it matters
On ice, for now.
Both agents are frozen today; the VM is not running. That is not a tidy success story, and I am not going to turn it into one. The fieldwork did what I needed it to do. I will pick it back up when I have time.
It showed where persistent agents earn their keep, gave me a working rule for where agents end and deterministic software begins, and sharpened how I evaluate agent products — including the ones I build at work.