Research index

Executive abstract

Grok Bot and Hermes Agent Bots are not two chatbots. They are two ways of packaging agency: a model proposing actions, tools executing them, a harness looping until a stop, and a human-facing glass where you name teammates and hand off work.

Confirmed (as of 2026-08-25): SpaceXAI (the xAI product surface) launched Grok Bot in early beta on 11 August 2026 as named AI teammates on a persistent cloud Linux VM, with connectors/MCP plus computer-use for apps that have no clean API.[1][2][3] Access rides Cursor / SuperGrok identity; there is no model picker; all of a member’s Bots share one computer (files, cookies, CLI credentials).[4][5][24] Nous Research’s Hermes Agent is an MIT-licensed self-hosted harness with 60+ built-ins, SKILL.md skills, cron, Kanban, and a messaging gateway; Bot Mode (public beta plugin 13 August, bundled default-on in v0.20.3 around 16 August) is a looking glass over Hermes profiles, not a new primitive.[8][21][22] A Nous Portal subscription is not Grok’s computer: it is inference + Tool Gateway (Firecrawl, FAL, OpenAI TTS, Browser Use, optional Modal sandbox) plus a preview always-on Hermes Cloud container billed to credits.[9][10][11][25]

Unverified / rumor: social claims that “each Grok Bot gets its own computer” (official docs say the opposite),[2][5] independent long-horizon reliability of either product, and Grok Bot SLA/disk retention beyond “durable workspace at /workspace.”[3]

The interesting story is mutual borrowing at insane pace, not a winner. Grok Bot copied the named-teammate roster, group chat, routines, skills, and “message it like a colleague” glass that open harnesses (Hermes, OpenClaw, Claude Code-class tools) had been iterating. Hermes Bot Mode copied the teammate roster UX two days after Grok Bot’s launch — while Hermes already had the underlying profiles, cron, and gateway. Grok did not copy Hermes’ open loop, model swap, or per-bot isolation. Hermes did not copy Grok’s account-scoped persistent VM with shared logins as the default product.

This comparison is also a stress test of the house Agentic AI Stack. The stack already distinguished harness (loop) from host (where commands run), but the public diagram still folds “computer” into Tools (execution environment / safe sandbox) and puts a host banner on Harness. Grok Bot makes the miss expensive: the scarce object is a durable runtime host with identity and disk, not a sandbox box and not the model.

Short answers

Question	Short answer
Same product class?	Same family (harness + bots looking glass). Different packaging: Grok = vendor-provided agent SKU on a rented VM; Hermes = open harness you operate, optionally rented via Portal/Cloud.
Persistent cloud computer?	Grok: yes, by default — one Linux VM per member, shared by all that member’s Bots.[3][5] Hermes without Portal: your machine / Docker / SSH. With Portal: Tool Gateway browser + optional Modal sandbox; Hermes Cloud preview = always-on agent container, not a furnished office VM.[9][11][13]
Who copied whom?	Same week: Grok Bot 11 Aug; Hermes Bot Mode beta 13 Aug; bundled ~16 Aug. UX convergence is real; host/isolation/model-swap divergence is the load-bearing difference.
Stack miss?	Treat runtime host / persistent computer as its own layer (or a first-class Harness service, not a Tools sandbox). Also: identity/subscription plane, shared-vs-isolated computer, teach-by-demonstration.

How to read this package

Each chapter answers one architecture question. A guiding question sits under every heading. Claims that move with the product are dated. X and video are signal; official docs are mechanism.

1. 01 — What each product actually is
2. 02 — Overlay on the Agentic AI Stack
3. 03 — Persistent computer vs Hermes hosts
4. 04 — Tools, skills, harness
5. 05 — Looking glass, bots, HITL
6. 06 — Nous Portal, SuperGrok, lock-in
7. 07 — Idea borrowing and release pace
8. 08 — Stack gaps, contrarian, outlook

Also: question ledger · references · podcast signal · illustrations