The AI Builder Podcast
How Dmitry Shapiro Built an Always-On Hermes AI Agent System
An agent becomes useful when the model is surrounded by a harness: tools, files, memory, triggers, permissions and monitoring. A tour of Dmitry Shapiro's always-on system.
What you’ll learn
- A language model becomes an agent when it is surrounded by a harness: tools, files, memory, triggers, permissions and monitoring.
- Shapiro's system runs named agents — Watson, Sherlock, Harry — across multiple machines, with Telegram and Linear as the control surfaces.
- GBrain is a searchable relationship graph built from Gmail, Slack, Zoom, LinkedIn, X and Calendar signals.
- Scheduled jobs, Tailscale networking and shared data spaces are what keep the system always-on rather than session-bound.
- Managing an agentic system is a different job from prompting a chatbot — you coordinate, monitor and delegate rather than type.
About this episode
A language model on its own is a conversation. It becomes a working agent when it is surrounded by a harness — tools, files, memory, triggers, permissions and monitoring — and that harness is what this episode is really about.
MindStudio CEO Dmitry Shapiro walks through his always-on system in unusual detail: agents named Watson, Sherlock and Harry running across multiple machines; Telegram and Linear as the control surfaces; Tailscale and shared data spaces underneath; scheduled jobs keeping everything moving; and GBrain, a searchable relationship graph built from Gmail, Slack, Zoom, LinkedIn, X and Calendar signals. Together they coordinate research, outreach, monitoring and personal workflows.
It is an architectural tour for experienced builders — the moving parts and the trade-offs, and why running an agentic system is a fundamentally different job from manually prompting a chatbot. Not a beginner's installation guide, and better for it.
Questions this episode answers
What does an AI agent harness include?
Tools, files, memory, triggers, permissions and monitoring. The language model supplies the reasoning; the harness is what turns it into a system that can act on its own.
What is GBrain?
GBrain is Shapiro's searchable relationship graph, built from Gmail, Slack, Zoom, LinkedIn, X and Calendar signals — the shared memory his agents draw on for research and outreach.
Is this episode for beginners?
No — it is an architectural tour for experienced builders. It shows the moving parts and trade-offs of an always-on agent system rather than a step-by-step installation.
