
Agentic Conversations (formally mlops.community) My AI Agent Bought Me A Car
26 snips
Nov 9, 2026 Sam Herring, an AI researcher and engineer at Nous Research, shares how Hermes Agent evolved from a productivity tool into an autonomous assistant that found him a retired parking-ticket car. They explore agents writing their own skills, recursive self-improvement through failure analysis, trace-to-environment training, the cost of overthinking, activation steering for smaller models, and the future of continual learning.
AI Snips
Chapters
Transcript
Episode notes
Reusable Skills Create Hermes’s Magic
- Life-admin benchmarks test bills, email, and other tasks outside conventional coding-agent evaluations.
- Hermes becomes more useful when it converts successful repeated workflows into reusable skills automatically.
Endless Terminals Trained Better Tool Use
- Sam moved from benchmark-focused RL to Endless Terminals, which procedurally generated increasingly difficult terminal tasks.
- The project trained models to improve through repeated, varied terminal-management challenges.
Hermes Could Hot-Swap Its Own Training
- An early Hermes could design an RL environment, train remotely through Tinker, download new weights, and hot-swap them into a local model.
- Sam compared the experience to Neo learning Kung Fu instantly.

