Video summary

The Ultimate Beginner’s Guide to Hermes Agent

Main summary

Key takeaways

Technology

Summary

The video is a beginner-to-advanced guide for “Hermes Agent”, positioning it as an AI “employee” / chief-of-staff that runs on a computer and can operate across tools, channels, and workflows—becoming more useful over time via self-improvement memory loops and reusable skills/workflows.

1) What Hermes Agent is (and why it’s different)

  • Hermes is described as a “harness”/runtime around an AI model (e.g., “GPT 5.5”, Claude Opus, etc.).
  • The core differentiator vs other agents/harnesses (examples mentioned: OpenClaw, Cloud Code, Codeex) is:
    • Hermes behaves like an AI employee that grows with you, not a one-off tool.
    • It uses a self-improvement loop where repeated successful work is stored as a reusable skill, so future tasks execute closer to your preferences automatically.
  • The video emphasizes Hermes becomes more reliable over time by:
    • improving/archiving skills strategically,
    • avoiding “bloated” or breaking agent behavior.

2) Fundamentals before setup: security, cost, and limitations

  • An analogy is used:
    • Model = engine
    • Hermes = car harness enabling computer control, channels, and memory
  • Claims:
    • Model-agnostic design
    • Local data: information stays on the user’s machine to reduce vendor lock-in
  • Key caution:
    • These agents are not fully autonomous decision makers
    • Users must:
      • limit permissions
      • verify outputs
      • add safeguards before high-risk actions (email, money movement, publishing)
  • Example risk mentioned:
    • an agent with email access allegedly deleted years of emails
    • takeaway: controlled access + verification is essential

3) Install & set up Hermes (two deployment paths)

  • Two deployment options:
    • VPS/cloud server
    • Local machine (e.g., Mac Mini / old PC)
  • The demo focuses on Mac setup.
  • “Buddy system” concept:
    • Use Codeex as a developer buddy/doctor to troubleshoot Hermes issues.
    • Suggested as “insurance” when setup breaks or remote help is needed.
    • Claim: Codeex can securely interact with apps from a phone even if the Mac is locked/off.

4) Model configuration and backups

Hermes desktop setup includes selecting:

  • Primary model
    • Recommended example: OpenAI / Codeex-branded plan using GPT 5.5 via subscription
  • Fallback/backup model via Newsportal
    • Described as a “backup brain” with access to top-tier models and audio models
  • Mentions:
    • “fast mode”
    • cost optimization via a cost-effective model mix

5) “Computer Use” integration (controlling your desktop)

Hermes “computer use” is presented as a capability for the agent to:

  • install required drivers/packages,
  • request permissions (including “KUA driver access” in subtitles),
  • then control the computer after access is granted—enabling hands-free task execution.

6) Multi-channel “gateways” (Telegram, Slack, iMessage, etc.)

The tutorial stresses connecting messaging channels first because they enable:

  • measuring work done while away,
  • tracking automation percentage,
  • treating Hermes like an “AI employee” and measuring “revenue per employee”.

Telegram

Hermes supports patterns such as:

  • a direct solo thread (Hermes replies),
  • a capture/ingest channel where Hermes can remember items without needing user responses.

Topic chats are demonstrated using:

  • BotFather,
  • bot tokens,
  • pairing codes,
  • admin permissions.

Topics enable multiple parallel threads (tabs) and show when Hermes is actively typing.

Slack

  • Configure using:
    • bot token
    • Slack app creation
  • Mentions:
    • socket mode
    • a “gateway restart” concept to ensure Hermes can talk to the channel
  • Supports team context and optionally whitelisting users.

iMessage

  • Uses Photon.codes to bridge iMessage into the Hermes gateway.
  • Phone messaging to Hermes is enabled after creating/applying credentials for a configured project number.

7) Identity + personality (“soul”, profiles, identity files)

Hermes personality is driven by identity files:

  • A “soul” file controls tone, style, and purpose
    • example behaviors: lower-cased, concise, “in-line”
  • An interview flow can generate user identity data (e.g., a “user.md” concept)

“Profiles” let users create multiple Hermes personalities/agents with different configurations:

  • roles,
  • models,
  • memory,
  • skills,
  • tool access.

8) Advanced memory: why multi-session agent memory is hard

The video frames memory as a complex systems problem:

  • Single-session memory
    • maintaining preferences and commitments within one long conversation
  • Knowledge updates
    • not treating old facts as still-current truth
  • Multi-session memory
    • connecting context across Telegram/Slack/other sessions
  • Temporal reasoning
    • determining what changed over time and what is now outdated

9) Honcho as a “shared brain” / memory router

Honcho is introduced as a memory layer that:

  • addresses “cold start” by routing/maintaining context across chats,
  • creates peer cards (a diacronic identity concept) representing relationships with different people/agents,
  • uses reasoning models (mentioned: “neuromancer”) to derive contradictions/implications.

Safety note:

  • Honcho is described as dangerous if given too much power,
  • but can be self-hosted for privacy.

Integration:

  • Hermes can use Honcho as its memory provider (via OAuth).

10) Agent management: Multica mission control + hooks

Multica is presented as an observability + orchestration layer:

  • manage multiple “runtimes” and workspaces,
  • track issues/projects/tasks handled by Hermes and other AI employees.

Hooks are used to create deterministic triggers, such as:

  • automatically updating Multica projects/issues when Hermes finishes tasks.

Demo theme:

  • Hermes acts like an AI project manager / chief of staff moving work across a board without manual saving.

11) Building blocks for real integrations: MCP, CLI, API, and connectors

Connectors communicate with external apps via interfaces like:

  • MCP
  • CLI
  • API

Examples:

  • Composio (MCP) as a gateway for many apps (Google Calendar/Gmail, Slack, HubSpot, Twitter/X, etc.)
  • Google Workspace CLI for Drive/Docs/Slides editing
  • cron jobs for recurring tasks

12) Cron jobs & “morning briefing” autopilot

A recurring automation is demonstrated that:

  • pulls calendar meetings,
  • summarizes urgent emails,
  • reads Slack team context,
  • finds AI trending topics (e.g., Twitter),
  • delivers a short practical briefing to Telegram.

Concepts used:

  • skills: turn repeated workflows (morning briefing format) into reusable units
  • agentic workflows: loops like research → curate → draft/review → deliver (not single prompts)

13) Build #2: automated animated website generation (“/slashgoal”)

Introduces /goal (slashgoal):

  • sets an outcome goal,
  • drives an iterative loop across turns until the goal is achieved.

Supporting tools/skills mentioned:

  • Higsfield AI (image/photo generation + deploy)
  • Firecrawl (scrape competitor sites)
  • here.now (deploy/hosting)

Outcome:

  • Hermes builds a premium animated landing page (graphics, copy, media elements, self-checking/refinement)
  • delivered as a deployed URL

14) Build #3: monetizing with “property marketing kits” (realtors)

Offer workflow:

  • scrape properties (Firecrawl),
  • generate marketing kits (upgraded visuals + room reconstructions),
  • compile sources and deliver assets/contact info for outreach.

Mentions:

  • a prebuilt skill in “cloud club”:
    • property marketing kit (described as a cloud-code skill)

Pricing (approximate ranges mentioned in subtitles):

  • about $250–$1,000 per listing

15) Build #4: content agent / growth automation

A content pipeline spans:

  • research trends/keywords,
  • scriptwriting,
  • filming/producing,
  • editing,
  • posting,
  • analyzing.

Emphasis includes:

  • carousels (Twitter/X-style and styled variants),
  • prebuilt skills for “Instagram thread carousel” styles,
  • a “content engine” automation driven by cron jobs and posting integrations.

Posting automation includes:

  • Composio
  • Blotato for scheduled posting calendars (noted: OAuth configuration needed)

“Content octopus” concept:

  • repurpose one YouTube video into shorts, articles, tweets, LinkedIn lead magnets/posts, carousels, etc.
  • uses a content repurpose skill

16) Video editing automation with skills

Demonstrates editing a raw talking-head clip via a two-phase approach:

  1. Cut video skill
    • removes filler/poor takes
  2. B-roll finder skill
    • finds B-roll and inserts it

Result:

  • a polished reel-like output.

17) Ecommerce ad factory + competitor ad inspiration

Ad creative generation approach:

  • use competitor research to guide messaging,
  • generate images/videos,
  • insert the creator into ads using a generated identity.

Mentions:

  • deploying/optimizing via an ads platform integration (example: Meta) at scale.

18) Community/customer support agent + inbox automation

A “school manager” / community manager agent:

  • checks new posts/comments via cron,
  • summarizes sentiment,
  • answers questions.

Integration:

  • Agentmail creates an agent mailbox:
    • create inbox,
    • provide Hermes credentials,
    • send/receive support emails,
    • manage support tickets.

19) Build: trading bot with self-improvement behavior

Trading bot framed as self-improving via Hermes self-improvement loops and skills.

Architecture:

  • broker/trading platform: co-invest
    • supports paper trading and switching to real trading
  • strategy/data source: QuiverQuant
    • congressional trading + strategies
  • connectors via MCP

Workflow demonstrated:

  • connect QuiverQuant MCP,
  • copy trade a person/strategy,
  • use cron jobs during trading hours,
  • generate weekly dashboards via here.now

20) Build: longevity/health coach with daily check-ins

Health workflow:

  • ingest blood panel PDF,
  • ingest sleep/workout sources (e.g., Oura/AW8 mentioned in subtitles),
  • agent produces a supplement plan,
  • daily reminders via cron,
  • voice memos to reduce friction and increase accountability.

21) “Jarvis mode” and a mission-control UI

A “Jarvis” command center includes:

  • calendar organization,
  • task agenda (“top three things”),
  • computer control,
  • voice output via 11 Labs (or local alternative, “voice box” mentioned).

Built by combining:

  • honcho for life context memory,
  • Composeio + Google MCP for calendar automation,
  • computer use,
  • voice,
  • HUD/UI built with design/coding orchestration (references include Claude + Fable + Codeex).

Theme: chief-of-staff UI + voice interaction.

22) Business + selling: pricing framework (productized services)

Go-to-market summary:

  • “Price the outcome, not hours.”
  • Recommended offer ladder (example pricing mentioned):
  1. AI employee install (non-technical): ~$1,500 setup (+ optional retainer)
  2. Executive assistant / autopilot (cron + automation): ~$500
  3. Animated website lead magnet/service: ~$250–$500
  4. Property marketing kits: ~$250 per listing (optionally scaled via autopilot)
  5. Content engine (carousels + posting): positioned as high leverage
  6. Content repurpose (weekly long-form → multi-platform): ~$1,000–$3,000/month
  7. Ad factory: creatives generation + scaling via ad platform integrations
  8. Community manager/customer support: ~$500 or per incident
  9. Trading bot + longevity coach: positioned as higher-ticket premium work
  10. Jarvis/Agent Club demo/performative product for closing calls

Advice emphasized:

  • demo and build first for yourself,
  • use a risk-reversal approach (e.g., ensure the client doesn’t bear most setup risk).

Main speakers/sources

  • Main speaker: the course creator/instructor (first-person throughout; name not provided in subtitles)
  • Referenced external source: Mark Cuban (motivational/business framing clip included)

Original video