Six AI agents that answer from the founder's own knowledge — and run overnight

A non-technical founder wanted an AI staff that works while he sleeps, with roughly thirty minutes of his own involvement: hand over the keys, test it, and then leave it alone. The demand was not for a smarter chatbot — it was for staff. Something that shows up in the morning with work already done.

The real problem only appeared after delivery. Autonomous systems fail silently. Nobody files a ticket when an agent stops working, because there is no user sitting in front of it.

We built six specialised agents — an orchestrator plus content, development, email, video and error-triage specialists — each running continuously and each reachable through its own chat bot, so the founder talks to whichever one he needs from his phone.

They self-delegate: the orchestrator can spawn the other five, with a hard cap on how many turns they may pass work between themselves so a loop can't run away. Every agent carries its own identity and memory files, so behaviour persists across restarts. Underneath sits a semantic memory layer — a vector index rebuilt every four hours over the founder's own sources, plus a structured knowledge graph — so the agents answer from his material rather than from general knowledge.

In practice they send a weekday morning briefing, triage the inbox, review pull requests on his repositories, and auto-fix errors reported by monitoring. Custom skills were written for the gaps the platform didn't cover, including a multi-model debate skill and an automated error-fix skill.

The engagement did not stop at delivery. It became a six-month operations relationship, which is where the harder work happened.