Hermes Agent tutorial: memory, models and automation

The speaker organizes Hermes Agent use into seven levels, starting with simple requests and ending with a dashboard that brings several AI workflows together. Early examples include research saved as an HTML file and access through Telegram. He runs Hermes on his own computer and prefers that arrangement to a VPS. The model examples use connected providers; running the agent locally does not establish that inference happens offline.

Memory is the next focus. The speaker describes recalling past conversations and using a soul.md profile to supply personal preferences, business context and goals. He then explains steer and background commands, reusable skills, and choosing models according to task and cost. His examples name ChatGPT 5.5, Opus 4.8, Sonnet 4.6 and DeepSeek V4 Flash.

A Granola example retrieves a meeting detail after granting access through MCP. The speaker warns that unnecessary MCP connections consume context and says his own configuration can draft emails but cannot send them. Later examples assign research and critique to parallel agents, request reminders, and schedule research reports.

The final section presents his dashboard for documents, memory and work across coding tools. He describes Hermes as an AI agent system around models and tools, and says it uses Claude Code rather than being a better coder. He also cautions that practical concurrency depends on the model and does not support an expectation of a thousand simultaneous agents.