Agent Wikis

wikis / Hermes / wiki / summaries / transcript-switching-to-hermes.md view as markdown

Transcript — I am Switching to Hermes Agent

type: summaryconfidence: highupdated: 2026-04-12hermes_version: v0.8.0sources: 1

Architectural deep-dive by creator Nick on why he is switching to Hermes Agent from other frameworks. The video centers on Hermes's trajectory capture + autonomous skill creation loop as the core architectural bet that differentiates it from stateless agent frameworks. The second half positions Hermes vs OpenClaw as two fundamentally different architectural bets (personal-learning depth vs platform-breadth) rather than competitors, and walks through a live install.

Key Points

  • Core architectural thesis: Most agent frameworks treat agents as stateless — ask, answer, forget. Hermes is built on the opposing assumption that an agent should learn from every interaction and compound value over time. This is an architectural choice, not a marketing claim.
  • Trajectory capture is the learning primitive: After every complex task, Hermes records the full trajectory — every API call, every tool invocation, every decision, in order. Most frameworks discard this; Hermes persists it and exports it in ShareGPT format for fine-tuning your own models. See self improvement loop.
  • Five-step learning loop:
    1. User issues a real multi-step task (e.g. "pull Stripe revenue, cross-reference HubSpot, analyze in Python, post a three-insight summary to leadership Slack")
    2. Hermes chains 40+ tools to execute (15–20 steps, first run can take up to ~5 minutes)
    3. Full trajectory is recorded
    4. Trajectory analyzer asks: "can this be packaged as a reusable function?" If yes, it writes a skill, tests it, and stores it at ~/.hermes/skills/ as plain code on disk
    5. Next similar request runs the skill (faster, cleaner, following the encoded SOP) and refines it if requirements shift
  • Skills are live code, not pre-recorded macros: they live in your filesystem, they improve with each execution, and they can be shared via agentskills.io (described as "NPM for agents"). Hermes ships ~40 bundled skills; the real value is the custom ones your usage generates. See skills system.
  • Memory with nudges: Beyond trajectories, Hermes tracks user-specific patterns — preferences, working style, what matters to you. Over time it proactively suggests skills ("you probably want a skill that does X instead of Y") which you approve or reject. See memory system and approval system.
  • Your Hermes becomes a different tool than someone else's: shaped by your workflows, preferences, patterns. This is what "self-improving" means concretely.
  • Six deployment backends: local, Docker, SSH (creator's personal choice on a cheap VPS), Singularity, Daytona, and Modal serverless (~$5/month because it hibernates when idle and only bills on message arrival). Framework does not lock you in. See deployment backends.
  • Positioning vs OpenClaw — different bets, not competitors:
    • OpenClaw: TypeScript, built by Peter Steinberger as a weekend project, went 0 → 300k+ GitHub stars in 4 months (beat React's 10-year record in 60 days), Steinberger joined OpenAI in Feb, moved to an independent foundation. ~13,000 skills on ClawHub, ~2M MAU, 24 platforms (WhatsApp, Telegram, iMessage via Blue Bubbles, Teams, Discord, Slack, Signal, Google Chat, Line, Matrix, IRC, WeChat, etc.). Purpose-built as a messaging gateway / single control plane across every channel in your life.
    • Hermes: Python (92.5% of codebase), built by nous research — an actual AI lab that trains the Hermes model family. ~19,000 GitHub stars, ~200 contributors, v0.6.0 quoted in transcript (the creator's install session predates v0.8.0). MIT licensed, production-ready.
    • OpenClaw does NOT learn — no trajectory capture, no RL pipeline, skills are human-maintained plugins, memory is bolt-on (LanceDB, lossless-claw) that you install and configure separately.
    • Hermes platforms: 12 (Telegram, Discord, Slack, WhatsApp, Signal, email, Home Assistant, DingTalk, Mattermost, Matrix, Feishu, WeCom) — less than OpenClaw's 24 but gap is narrower than most realize. Cross-platform conversation continuity is a recurring reviewer callout (start on CLI, notify on Telegram, resume on Discord without context loss).
    • OpenClaw backends: mostly Node.js process or Docker. Hermes: six including Modal serverless and Singularity/HPC.
    • hermes claw migrate command: one command pulls your SOUL.md, memories, skills, API keys, and messaging configs from OpenClaw. Framed as a deliberate strategic move by Nous — "keep OpenClaw if it works, try Hermes alongside it." See migration from openclaw.
  • Recommendation framework:
    • Choose OpenClaw if you need an operations + platform play: agent across 5+ business channels from day one, massive tutorial ecosystem, customer-facing tasks, find-a-tutorial-at-2am community.
    • Choose Hermes if you want a personal agent + research play: compounds knowledge over time, gets faster the more you use it, exports training data for your own models, you're comfortable running on a server.
  • Honest limitations called out:
    • Windows not supported — WSL2 required, full stop. Experimental PRs landing for path/PTY handling but README says unsupported. Enterprise plan accordingly.
    • Learning loop only fires on complex tasks — simple tasks generate no skills. LLM-generated skills aren't guaranteed to work; you'll adjust them.
    • Local backend security surface — agent runs terminal as you; dangerous command checks exist but docs recommend Docker or Modal as the security boundary for production.
    • Stuck agent loops — ignore-then-loop behavior recurs in different forms; v0.4.0 changelog had multiple fixes; improving but not solved.
    • Speed — Hermes trades speed for learning. CrewAI and LangGraph beat it on raw orchestration throughput. If your use case is "run this pipeline as fast as possible," Hermes is the wrong tool.
  • Install experience: single curl bash command runs the installer, handles Python, Node, ripgrep, ffmpeg; ~60 seconds. Setup wizard prompts for provider (Anthropic/Claude subscription and ChatGPT plans both supported — unlike "OpenClaude" which restricts this), TTS, backend choice, sudo support, progress display, activity mode, session reset policy, platform adapters (Telegram setup is just BotFather → paste token), skill enablement, RL training opt-in.
  • Running from phone: hermes gateway starts the adapter; connect via Telegram bot, get the same agent / same memory / same skills accessible from anywhere.
  • VPS recommendation: creator uses a VPS (Hetzner, Hostinger style) so the machine isn't tied to his laptop. Offers to make a dedicated VPS setup video.
  • Strategic takeaway: The question is no longer "should we use agents" but "what kind of agents should we build." Hermes represents one answer — agents that learn — and architecturally that's surprisingly hard. Nous Research building this from the ground up signals they believe the future of agents is in adaptation/learning depth, not breadth/generality.

Relevant Concepts

Source Metadata

  • Type: YouTube video transcript (auto-captioned, lightly cleaned)
  • Title: "I am Switching to Hermes Agent"
  • Speaker: Nick (channel creator; states he has driven $5M+ in bottom-line AI revenue across clients and trained hundreds of entrepreneurs)
  • Approx date: Recorded against Hermes v0.6.0 (creator references v0.6.0 on-screen); pre-dates v0.7.0 and v0.8.0. Logged into KB on 2026-04-12.
  • Source file: raw/transcript-switching-to-hermes.txt
  • Cross-references to Hermes doc surface: curl installer, hermes launcher, hermes gateway, hermes claw migrate, ~/.hermes/skills/