Aram Shakhbandaryan

Agent operations

Hermes is an open-source agent framework from Nous Research. I did not write it. What is mine is the setup running on it: a private VPS that handles the recurring work around Decato and around me, driven from Telegram, one thread per subject, with memory that persists across months.

What runs

Twenty-two scheduled jobs in production, in three groups. Growth: nine jobs on Decato — Search Console pulls, an indexation watchdog, technical checks, a content backlog, a weekly operations report, lead digests and conversion alerts. Money: the weekly investor update draft, a sweep of stale follow-ups, a daily finance check-in. Upkeep: a nightly backup and a guarded weekly update of the system itself.

How I use it

Every subject is a Telegram thread and the thread title is the context: one for Decato's search traffic, one for leads, one for the investor update, one for finance. I write to it in Russian, it answers in Russian, and the files land in the thread. Bigger work runs as an orchestrator: I write the spec, sub-agents execute, I review, it merges. Rule-based triggers where speed matters; a model where context does.

The job I use most is the coach. Every evening it asks five questions — what shipped, what got added mid-day, what I avoided, how I am, what comes first tomorrow — I answer in three minutes of audio, and it comes back with the name of the pattern and one action. It has read seven years of my own notes, so it recognises my loops faster than I do. Two weeks in, scope creep was visible the same day instead of three days later, and I was getting four to six hours of focused work a day.

Teaching it

The first session in Yerevan drew thirty people and five volunteer mentors. AI9 Startup Campus hosted the next in May 2026, with attendees required to arrive with a server provisioned and SSH working.

Standing beside a projected slide reading "What Do You Want to Build?" in front of a room of people working on laptops
AI9 Startup Campus, Yerevan, May 2026.