NousResearch co-founder says Hermes Agent works better with memory, skills and separate evaluation
alexcovo_eth · x · 2026-08-04
Six takeaways from a NousResearch co-founder on getting more out of Hermes Agent
The post summarizes a podcast-style conversation with NousResearch co-founder Karan about Hermes Agent:
- The “personal” part of a personal agent comes from memory and the skills you build with it, not the underlying model.
- Karan says he barely notices when he swaps models because Hermes already carries so much context about him.
- He recommends trying different personalities via /personality, such as teacher, concise, kawaii, or pirate.
- He suggests using separate agents for doing the work and evaluating it, because a fresh evaluator is better at catching errors, weak assumptions, and missing evidence.
- He frames “you’re absolutely right” as reward hacking: the model is optimizing for approval, not truth.
- The thread also points to Hermes as a self-improving agent and mentions AI-assisted game modding.
More from coding & agent
- OpenCode Go says it processed 6T tokens in a single day, led by DeepSeek models — ycombinator · 2026-08-04
- Google sunsets consumer Gemini Code Assist on GitHub and ships a broken uninstall link — keunwoochoi · 2026-08-04
- The $2,000 repair estimate loses to the unmarked shop everyone already knows — bennash · 2026-08-04
- A Rust-style `#[predict]` joke imagines an API that always answers correctly — krypticmouse · 2026-08-04
- Defining "Listening" in Agentic Frameworks: Continuous Audio Propagation — juberti · 2026-08-04
- Codex uses Blender and Unity to build an otter-mech game with full assets — emollick · 2026-08-04