Narrative + technical walkthrough
How I Built a 6-Agent AI Operating System
Perspective: Builder / Practitioner
The first version did not feel like an operating system. It felt like a collection of clever scripts that needed me to remember what to run, when to run it, and how to recover when something failed. That was the problem. If the human is still the scheduler, the debugger, and the memory layer, the system is not really autonomous.
The architecture that survived
The system settled into six practical roles: intake, research, execution, verification, memory, and notification. Each role had a job, a boundary, and a failure mode. That mattered more than the model choice. A powerful model without routing, state, and verification still behaves like an assistant waiting for instructions.
- Intake: normalize requests and decide whether the job belongs in chat, cron, or Kanban.
- Research: gather outside facts, documents, and prior context before writing or acting.
- Execution: perform the actual file, browser, terminal, or API work.
- Verification: prove the output works instead of assuming it works.
- Memory: preserve stable preferences and environment facts without turning every session into a diary.
- Notification: return the result through the right channel with proof.
Failures that changed the design
The biggest failures were not model failures. They were coordination failures. A worker would finish but not preserve the artifact. A browser cache would make a new page look old. A process would succeed in GitHub but remain stale on the custom domain. Those failures forced the system to treat verification as a first-class agent responsibility.
The tradeoff
More agents create more surface area. They also reduce human cognitive load when each agent has a narrow, verifiable job. The lesson was not “add more agents.” The lesson was “separate judgment, execution, and verification.”
What I would build again
I would keep the operating-system pattern: persistent memory, scheduled work, parallel execution, local verification, and channel-specific delivery. It turns AI from a chat window into business infrastructure.