Hermes Agent System

Our client was managing eight revenue-generating products simultaneously — each with its own content pipeline, social media presence, competitive landscape, and growth strategy. The founder was spending 60-70% of their time on operational tasks: scanning Reddit and Hacker News for market signals, scheduling social posts across platforms, monitoring an autonomous trading system, and context-switching between products dozens of times a day. The core issue was not a lack of tools. It was that no single tool could hold the full context of a multi-product business and act on it without constant human direction. They needed a system that could think across products, not just execute within one.

We designed and built Hermes, a multi-agent AI orchestration platform running on a dedicated VPS we named Olympus. Rather than one monolithic assistant, we created four specialized AI personas, each responsible for a distinct operational domain: - **Midas** — An autonomous CRO agent that handles revenue optimization, generates content, and runs social media campaigns across all eight products. It has direct access to a self-hosted Postiz instance for publishing and uses Playwright for browser-based automation when APIs are not available. - **Oracle** — An intelligence and research agent that continuously monitors Reddit, Hacker News, Twitter, Product Hunt, and Google Trends, surfacing relevant signals and competitor movements. - **Plutus** — A trading monitor that watches over an autonomous trading system (WealthPilot), manages risk thresholds, and holds kill-switch authority to halt trading if conditions deteriorate. - **Kuber** — A personal assistant that serves as the human interface, routing tasks between agents via Telegram based on intent. A content approval dashboard (VIRALIS) sits between the agents and the outside world, ensuring nothing gets published without a human check when required.

The system runs on a Node.js dispatcher that spawns Claude Code agents per persona, each with session persistence, dedicated workspace files, its own system prompt, and a scoped set of allowed tools. This means Midas cannot access trading controls and Plutus cannot publish social content — separation of concerns enforced at the agent level, not just by convention. A cron-based scheduler triggers routine tasks (market scans, content generation cycles, trading health checks), while a Discord bot and Telegram bots handle ad-hoc requests. We built a Mission Control dashboard in Next.js backed by SQLite to give the client a single view of what every agent is doing, what is queued, and what needs approval. We chose self-hosting over managed AI platforms deliberately. Running on a dedicated VPS gave us full control over agent memory, session state, and cost — critical when agents are running continuously, not just responding to one-off prompts.

The founder reclaimed roughly 40 hours per month of operational time. Oracle catches market signals that previously went unnoticed for days. Midas maintains a consistent content cadence across all eight products without the founder writing or scheduling a single post manually. Plutus has intervened twice on trading anomalies that would have required the founder to be watching a dashboard at 3 AM. The system manages revenue strategy across all products simultaneously — something that was simply not possible with one person and a collection of disconnected SaaS tools.