Jonathon's AI Wiki

Hermes Personal-Agent Briefing (Dmitry Shapiro, MindStudio)

7 min read

Source: raw/The_Hermes_Agent_Briefing_-_What_It_Is_Why_It_Matters_and_How_to_Use_It_Today.md

Creator: Dmitry Shapiro (CEO, MindStudio; ex-Google product, ex-MySpace CTO, founder of Veoh + Iconic Systems) ^[Auto-captioned “Dimmitri Shapiro”; normalized to the verifiable MindStudio CEO whose bio matches the source exactly.] URL: https://www.youtube.com/watch?v=yzlvDnxvi1I Platform: YouTube (live webinar/seminar, co-hosted with Josh; bootcamp taught with Dr. Marily Nika)

A practitioner briefing framed as “what it is / why it matters / how to use it” rather than a step-by-step install. Shapiro runs Hermes in production as a personal agent named Watson (always-on, old MacBook Pro, 18 cron jobs) and uses it as the lens to explain the whole “personal agent” category. The reusable value is the harness mental model, a concrete cost-control pattern, a long-term-memory layer (GBrain), two customization dimensions, and two metric-backed business case studies — most of which aren’t covered by the existing intro/course articles.

Key Takeaways

  • A “personal agent” = a harness + a model + your credentials. A harness is an app that wraps a model with tools so it can drive your computer — read/write/delete files, run shell commands, drive a browser, write code, self-improve, manage long-term memory and context compaction. Shapiro groups OpenClaw / NemoClaw / NanoClaw / ZeroClaw / Hermes under the umbrella term “claws.” He cites Jensen Huang’s^[Auto-captioned “Jensen Kuang.”] line that “OpenClaw is the new computer” as the thesis: you now buy a computer and give it a harness + model, like buying a self-driving car instead of driving it yourself.
  • Shapiro’s current pick is Hermes (his opinion; “by far the best… all of that is bound to change”), run as the Hermes Desktop app — which he calls the right way to run it.
  • Activity-dependent model selection is the main cost lever. Don’t run one premium model for everything. His Anthropic setup: Sonnet as default, auto-switch to Opus only for real coding/tool-building then switch back, Haiku for cheap NLP (parsing, small summarization). Copy-paste instruction to the agent: “implement activity-dependent model selection.”
  • Build-your-own is viable: point a coding agent (Claude Code / Codex / Antigravity) at the Hermes or OpenClaw repo and say “build me a lighter custom personal agent” — you become the PM of your own harness.
  • Connect to a model two ways: OAuth to your subscription (cheaper, highly-discounted tokens, but rate-limited) or the API (flexible pay-as-you-go). He recommends subscription/OAuth to start.

The Two Customization Dimensions

Distinctive framing for configuring an agent (both are “it depends” — experiment):

  • Digital-you vs. digital-assistant. Decided largely by credentials: give it your own logins (it acts as you) or create a fresh assistant identity (separate Gmail, defined role). Shapiro runs Watson as himself; the only tell is a fox icon Hermes appends to its own LinkedIn/Twitter posts.
  • Indoctrination vs. acculturation. Indoctrinate = prescriptive guardrails, “just do the work, don’t give advice.” Acculturate = teach it your culture/norms and let it act proactively (it researches, returns a plan, asks to execute). Acculturation is the path to a chief-of-staff-style agent.

GBrain — Long-Term Memory Layer

A third-party stack (attributed to Gary Tan) Shapiro layers on top of Hermes; no prior wiki coverage:

  • Install: paste one command into the agent; needs an OpenAI API key (uses an OpenAI endpoint for vector embeddings).
  • What it does: ingests email / calendar / Zoom / Slack / Notion / Twitter / LinkedIn → maps a graph → vectorizes everything for semantic search. Stores files in Obsidian (free) and renders the graph there.
  • Smart contacts / entities: daily ingest + search-API enrichment keeps a self-updating CRM fresh (job changes, company moves, what people post about) — replacing decaying manual CRM entry.
  • “Dreaming”: a nightly process that refactors/enriches the stored data so it stays tidy and queryable.

Setup Essentials & Housekeeping Crons

  • Configure a search API early. Agents search constantly to extend/fact-check model knowledge (driving a real browser to Google is “dumb / slow — made for humans”). Recommends you.com (~$200 credits; called a must for financial data).
  • Prioritize housekeeping crons first (his stated build order): search API → GBrain → cost/efficiency monitoring → model-drift monitoring (catch when non-deterministic behavior silently switches models or runs away) → nightly security audit + patch. A nightly “inefficiency scan” self-analyzes the (constantly-changing) setup for poor architecture and runaway spend, then proposes fixes. He runs 18 crons total.
  • Warning: this is a serious time sink (“don’t get divorced”).

Two Business Case Studies (claims, with metrics)

  • MindStudio’s agent-built SEO blog. Per an Ahrefs screenshot he showed: organic traffic ~3–4K → 140K+/month in ~3 months. ^[ambiguous] Pipeline = 3 subsystems: (1) every 24h, watch 11 hand-picked AI-news YouTube channels, pull transcripts, rank topics as a set (YouTubers as a proxy for what people want to know); (2) deep-research each topic, generate multiple angles, auto-write 20–60 SEO posts/day; (3) a sub-agent generates hundreds of ads for their new product “Remy.” Built by a non-technical team member — no developer or designer involved. Articles get cited by AI models and rank in search.
  • Investor-outreach pipeline. A ~1,200-investor spreadsheet (source-of-truth tracker) the agent built and enriches via deep research — thesis-match, prior investments, social/news signal → proposed hook + custom cold-email draft. He keeps a human-send gate (drafts only), then it monitors his inbox for replies and queues follow-ups from new angles.
  • GrillMe (extreme acculturation): an interview pattern where the agent exhaustively questions you and writes markdown context files. Example shown: an anonymized file set for a level-5 senior PM on Meta’s infra team — role, company culture, procedures, what it can do solo vs. escalate, key peers. This source-of-truth gets compressed into context so the agent runs “as” that role.

Q&A Highlights

  • Security / separate device: good idea for beginners. A cheap always-on box works (Mac minis sold out due to claw demand; NanoClaw/ZeroClaw run on a Raspberry Pi; or a VPS like Hostinger). Shapiro now runs Hermes on his primary machine with full credentials — “fortune favors the bold.”
  • Multiple instances: yes — one Hermes acting as many sub-agents, multiple Hermes processes, or a fleet of Mac minis / VPSes. Cites the Karpathy / Boris Cherny^[Auto-captioned “Boris Churnney.”] point that humans should write loops that prompt the coding agent, not prompt it directly (meta-orchestration to scale).
  • OpenClaw vs. Hermes: OpenClaw (built by Peter Steinberger; Shapiro’s claim that it was acquired by OpenAI) was first-to-big with a huge community (Claw Hub, skills sites) but, in ~2 dozen of his installs, “30–40% a disaster” — janky/bloated. Hermes has impressed him by not being that. Or build your own.

Try It

  • Add activity-dependent model selection to an existing Hermes/claw config: instruct it to default to a mid-tier model, escalate to a top-tier model only for coding/tool-building, and drop to a cheap model for parsing/summarization.
  • Stand up housekeeping crons before feature crons: a nightly cost/inefficiency scan, a model-drift check, and a security-audit-and-patch job.
  • If you want a self-updating CRM, evaluate a long-term-memory + graph layer (GBrain-style: ingest your comms → vectorize → Obsidian graph → nightly enrichment) and wire a search API (e.g., you.com) for entity enrichment.
  • Pick a customization stance deliberately: digital-you vs. assistant (credentials) and indoctrination vs. acculturation (guardrails vs. proactivity).
  • To prototype a custom harness, point Claude Code / Codex at the Hermes or OpenClaw repo and ask it to build a slimmer, made-for-you agent.

Open Questions

  • Unverified-from-source claims (single practitioner, some shown only via screenshots): the 3–4K→140K traffic figure, “20–60 articles/day,” “~1,200 investors,” and the OpenClaw-acquired-by-OpenAI assertion. Treat as Shapiro’s claims pending corroboration.
  • GBrain’s exact provenance, license, repo, and ongoing cost (beyond the OpenAI-embeddings key) aren’t specified in the source.
  • The shared resource doc / use-case list and the Meta-PM GrillMe example file referenced in the talk weren’t captured in this transcript.

Graph View

Backlinks