Key Takeaways
- Modal serverless execution means compute costs only accrue when the agent is actively working — near-zero idle cost.
- Custom skills built through repeated use compound: each research task makes the next one faster.
- Telegram as the interface means research requests can be made from anywhere — the agent meets you on mobile.
The Architecture
The stack separates concerns cleanly: a $5 VPS handles always-on availability and Telegram interface. Modal handles serverless compute for research tasks — cheap when active, zero cost when idle. Telegram provides the conversational interface accessible from any device.
- $5 VPS: always-on availability, lightweight process management
- Modal serverless: compute that scales to zero when not in use
- Telegram: conversational interface accessible from any device
- Custom skills: accumulated research procedures from past sessions
The 40% Improvement
The 40% speed improvement over a fresh-session agent comes from the accumulated custom skills. Each research task adds to the skill library — refining source selection, output structure, and filtering criteria. A fresh agent starts every task from zero. An agent with accumulated skills starts from a higher baseline.
“40% faster on research tasks”
The Consulting Application
An additional pattern worth noting: for consulting work, the VPS can be placed inside a client's VPC, giving the agent access to internal systems while maintaining the Telegram interface. The same agent architecture serves both personal research and client-facing work with a network configuration change.
Story sourced from the official Nous Research Hermes user-stories page. Original author: Krzysztof Słomka.