Glass & HITL

Users meet a glass, not a stack

❓ What does a human actually see, and which HITL primitives are enforced rather than hoped?

A looking glass is the surface for goals, step visibility, gates, and outputs. Behind it: a harness.

Grok’s glass

❓ Why does Grok Bot feel like iMessage with a coworker?

• Desktop macOS/Windows + iOS; same thread on both.[4]
• Sidebar of named Bots; pin/hide; up to 50 Bots + group chats combined.[20]
• Create: name, job, description, avatar — operational role text, not a workflow builder.[4][20]
• Agent Computer preview of clicks; takeover for human-only steps.[3]
• @ for Bots, groups, routines, connectors; / for skills.[18]
• Group chat: Bots pass work and pull you for judgment.[1]
• Duplicate copies profile/skills/routines/avatar not conversation memory.[20]

There is no Linux glass. The computer is Linux; you watch it from Mac/Windows/iOS.[5]

Identity gravity: Cursor SSO. Legacy Privacy Mode blocks the product entirely.[4][5]

Hermes’ glasses

❓ If Bot Mode is optional UI, where do Hermes bots live when Desktop is off?

Glass	Role
CLI / TUI	Harness-native. hermes -p name chat is the Bot.[21]
Desktop Sessions	Chat list, classic agent.
Desktop Bot Mode	Roster, forever-chat, routines pane, groups, @mentions.[21]
Gateway	Telegram, Discord, Slack, WhatsApp, Teams, … 20+.[8] Pairing/auth for DMs.
Hermes Cloud	Preview: same agent, many channels, memory not tied to device.[11]

Bot Mode specifics that Grok also has, almost phrase-for-phrase: New Agent (name/title/description), Advanced clone, hide without deleting, routines beside the Bot, group rooms, @mention handoff, “needs you” escalation.[21]

Hermes-only glass details:

• Forever-chat: /new inside canonical Bot Chat becomes /compact so the relationship does not fork.[21]
• Bots are files on disk; CLI parity table is complete.[21]
• Rooms span machines; @name-device; hermes peer dm without Desktop in the loop.[21]
• Turning the plugin off leaves profiles, sessions, cron.[21]

HITL primitives

❓ Which of the five gates exist as product, not as a sentence in a system prompt?

House primitives: (1) gate on action class (2) pause / durable blocked state (3) context packet (4) recorded decision (5) resume or cancel.

Primitive	Grok Bot	Hermes
Gate on class	Auto Review rules + local-computer Never/Ask/Always; team ceiling “coming soon.”[5][19]	approvals.mode; toolset disable; YOLO as explicit bypass
Pause / durable	Routines pause after long absence; computer recover/reset; takeover blocks.[18][3]	Kanban blocked; approval prompts; cron
Context packet	Approval card shows operation + inputs.[19]	Approval UI / gateway /approve /deny
Recorded decision	Desktop/iPhone allow/deny; org audit view coming.[5]	Session transcripts you own; Kanban comments
Resume	Return control; “Always allow” matching rule.[19]	Continue loop after approval

HITL ≠ “humans use the chat.” Grok’s best HITL is takeover of the VM for 2FA — a genuine primitive. Hermes’ best HITL is you own the log plus Kanban block. Both still lean on prose boundaries (“never send without approval”) for a lot of the blast radius.

Roles

❓ Who builds the bot, who writes the skill, who reviews tools?

Role	Grok	Hermes
End user	Creates Bots, teaches tasks, approves	Same, plus CLI
Citizen builder	Teammate roster, no workflow canvas (by design)[1]	Bot Mode + SOUL.md
Platform	Cursor admins: MCP policy, Cloud Agents toggle, privacy mode[5]	You / your IT: backends, toolsets, Portal
Domain expert	Bot description + skills	SKILL.md authors
Vendor	SpaceXAI/Cursor loop and models	Nous optional Portal/Cloud; MIT core

Building a Bot ≠ training a foundation model. Grok will not let you pick the model anyway.[5]

Looking-glass lock-in

❓ If the critical path is single-homed on one glass, what fails at 2 a.m.?

Grok: if Cursor identity or the desktop/iOS app has an incident, you have no second glass except waiting. The VM may still run routines.

Hermes: if Desktop dies, CLI and gateway remain. If Portal dies, BYOK remains. If your laptop dies and you only ran local, work dies — unless you already operated SSH/Cloud.

That is the operate-vs-rent split again, now at the glass.