Hermes Agent
Agents generate Skills. The next harness layer.
Nous Research's Hermes Agent implements a closed learning loop where agents generate and improve their own skills. The old paradigm: humans write skills, agents execute. Hermes: agents do complex work, auto-generate reusable skills, and self-improve with each repetition. The author of skill generation has moved from human to AI.
- 01Agents make skills themselves and fix them as they run
- 02FTS5-based session search for past-conversation recall
- 03Honcho for accumulating user models
- 0415 channels unified behind one gateway (Telegram/Slack/Discord/…)
- 056 backends: local, Docker, SSH, Daytona, …
- 06Modal + Daytona serverless hibernation — idle cost 0
- 07Only agent with closed learning loop built-in
- 08Who writes skills shifted: human → agent, human sets direction only
Agent runs tasks, creates reusable skills, and improves them on every reuse — a self-tuning feedback loop.
Reusable procedures generated automatically by the agent after complex tasks. Previously written by humans.
FTS5-based retrieval of past conversations and context to improve agent decisions.
“Agents make skills themselves and fix them as they run”
— Goobong Jeong
“The big shift is who authors the skills”
— Goobong Jeong