Kihagyás

Hermes Agent — Agents That Grow With You (Practical AI, 2026-05-21)

Source: Practical AI Podcast Guest: Jeffrey Quesnelle (CTO & co-founder, Nous Research) Hosts: Daniel Whitenack, Chris Benson Date: 2026-05-21

Overview

Jeffrey Quesnelle, CTO of Nous Research, discusses the evolution of Nous from a Discord community to a leading open source AI company, the philosophy and architecture behind Hermes Agent (the #1 open source agent repo on GitHub), the geopolitics of open source AI, and practical advice for agent usage.

Key Topics

Nous Research Evolution

  • Started as Discord community of independent AI researchers ("loose collaboration of the homies")
  • Formed company ~2 years ago to let open source researchers work full-time
  • North Star: keep AI open, accessible to as many people as possible
  • Pivoted through: academic research → Hermes fine-tunes → distributed training research → Hermes Agent

Hermes Agent Architecture

  • Built by "Technium" — a non-developer who used AI tools to create the #1 GitHub open source agent repo
  • Core philosophy: "The agent that gets better the more you use it"
  • Skill system: agent autonomously observes patterns, creates reusable skills. Example: Las Vegas restaurant booking — took 30-45 min first time (captchas, API discovery), instant the second time
  • Hierarchical memory: notes about user, past sessions, communication preferences — all activated at LLM's discretion
  • Minimal hard-coded features: only code execution, web browsing. Everything else emergent through prompts
  • Nous employs "LLM whisperers" who craft prompts to encourage self-reflection
  • Local-first: runs with local models as first-class citizen; hosted option available

Model vs. Harness

  • Analogy: model = brain, harness = body
  • IQ 150 in wheelchair limited vs IQ 95 with athlete's body — symbiotic relationship
  • The harness touches reality; models only output tokens
  • Future: outcome-based payment — pay for jobs done, not tokens

Open Source AI Geopolitics

  • Western retreat: Llama 4 failure ($200-400M paperweight) killed Meta's enthusiasm; Western open source models largely disappeared
  • Chinese surge: DeepSeek near-SOTA shocked geopolitics; Chinese companies use open source as growth hack — release model → instant visibility
  • NVIDIA's $20B commitment: Jensen Huang at GTC committed to training Western-provenance open source models — NVIDIA uniquely aligned: everything runs on their chips regardless ("the house always wins")

Practical Agent Usage Advice

  • Automate: tasks requiring infinite patience but zero creativity
  • Don't think in human roles (CEO agent, CFO agent) — think operations
  • Don't tell the agent HOW — describe outcomes and success criteria
  • "Get out of the way of the model" — long-horizon planning is improving
  • Be explicit: unspoken assumptions persist because AIs lack shared human experience
  • "Explain it like I'm an alien" — because the AI is one
  • AI limitations: no aesthetic judgment (no lived human experience), regresses to the mean (not truly novel)

Nous Dogfooding

  • 99.99% of Hermes Agent code written by Hermes Agent
  • Multiple agents deployed internally: backend debugging via MCP, read-only infrastructure access
  • One engineer showed it once → now support staff uses the same agent
  • Emergent knowledge transfer across organization

Personal

  • Quesnelle had a baby 5 days ago — contemplating what world she'll grow up in
  • Mission: human-centric AI — "better today than yesterday, better tomorrow than today"

Notable Quotes

  • "The agent that gets better the more you use it"
  • "Model is the brain, the harness is your body"
  • "Infinite patience, but very little creativity"
  • "Show me incentive, I'll show your outcome"
  • "There's never been a time more where a single person's leverage can be maximized"
  • "Explain it to me like I'm an alien — because that's really what the AI is"
  • "If you have the vision and the drive and the care, this technology can make you a thousand X"

Tags

ai-agents, open-source-ai, nous-research, hermes-agent, agent-architecture, llm, agent-harness, nvidia, deepseek, geopolitics

Vissza a tetejére