Stack overlay

The house map, used as a measuring stick

❓ If we drop both products onto Model / Persistent State / Tools / Skills / Harness / Multi-agent / Looking Glass, which boxes are full, empty, or mislabeled by the vendors?

The house stack (bottom → top) is:

Layer	Job
Model	Reason, plan, choose next action
Persistent State	Pre-existing files, databases, enterprise systems of record
Tools	Side effects: built-ins, CLI, MCP, connectors, one execution environment / safe sandbox
Agent Skills	Versioned how-to (SKILL.md, progressive disclosure) — in-house know-how you cannot buy
Agent Harness	Loop, dispatch, context, memory, approvals, sandbox policy, budgets, tracing
Multi-agent orchestration	Board / graph / router / sub-agent / A2A — bonus level
Looking Glass	Messaging apps, native desktop, HITL

Core mechanism (hard to vary): Agency is closed-loop control over side effects. Model proposes → tools execute → skills bias procedure → harness controls the loop → HITL gates high-impact writes → looking glass surfaces the loop → orchestration only when multi-role ownership must survive crash and handoff.

Neither vendor’s marketing uses these words. Mapping them is the point.

Layer-by-layer fill

❓ Where does each product actually put work, versus where the brochure puts it?

Layer	Grok Bot	Hermes Agent Bots	Overlay note
Model	Fixed routed set; no picker for members or admins.[5] Grok-family plus failover subprocessors.	Swappable: Portal catalog (300+ including Grok, Claude, GPT, Gemini…), OpenRouter, Codex OAuth, local.[8][9] Different Bots can pin different models.[21]	Grok locks cognition inside the SKU. Hermes treats LLMs as the swappable band.
Persistent State	Real apps (Salesforce, Gmail, Slack…) via browser/connectors. /workspace on the VM is agent workspace, not your SoR.[3]	Your repo, DB, and enterprise systems via CLI/MCP. Profile files under ~/.hermes/ are harness state, not SoR.	Both can touch SoR. Neither should become SoR. Grok’s shared VM is the easy place to accidentally store secrets.
Tools	Plugins (connectors), MCP, browser computer-use, terminal on the VM, optional local computer with approval.[3][5][19]	Built-ins (60+), CLI, MCP, Tool Gateway (search/image/TTS/cloud browser), seven terminal backends.[8][13][25]	Grok’s distinctive tool is computer use on a durable desktop. Hermes’ distinctive tool is CLI + local/high-bandwidth loop.
Skills	Saved instruction packs; / to invoke; teach-by-demo (~10 min recording); plugins marketplace; per-Bot enable of private skills.[18]	agentskills.io SKILL.md + scripts/refs; skill_manage; curator; hub.[8][26]	Grok skills look like prompted procedures + demo. Hermes skills are versioned packages with progressive disclosure. Not the same LTC depth.
Harness	Managed loop inside the Grok Bot app/cloud. Auto Review, approvals, budgets via Cursor usage. Not customer-inspectable.	run_conversation() loop, compression, approvals, tracing, checkpoints — the product.	Grok hides the harness. Hermes is the harness.
Multi-agent	Many named Bots, group chats, independent messaging, “chief of staff” pattern.[1][2] Not a crash-durable Kanban board. Shared VM, not separate workers.	delegate_task (RPC, dies with parent); Kanban (durable SQLite board + profile workers); Bot Mode group rooms (2–6 Bots, 3 rounds).[8][21][28]	Grok’s multi-bot is collaboration UX. Hermes Kanban is the bonus-level board. Do not equate them.
Looking Glass	Native desktop (macOS/Windows) + iOS; message-thread UX; Agent Computer live view + takeover.[4][3]	CLI, Desktop (Sessions + Bots tabs), 20+ gateway glasses.[8]	Grok: one polished glass. Hermes: many glasses, one harness.

Category errors this overlay forbids

❓ Which vendor phrases would make an architect buy the wrong layer?

1. “Bots” ≠ multi-agent orchestration. A roster of named loops is a looking glass. A board with crash reclaim is orchestration.
2. MCP ≠ the computer. Grok’s own pitch is that computer use exists because MCP/API is missing.[1]
3. Shared VM ≠ per-bot sandbox. Screens are separate; security is not.[2][3]
4. Harness ≠ host. Grok’s host is the Linux VM. Hermes’ harness can sit on local, Docker, SSH, Modal, Daytona, Vercel sandbox, Singularity.[13]
5. Skills ≠ memory. Grok docs correctly separate Bot memory from /workspace files and from SoR.[20]
6. Vendor agent ≠ a new stack layer. Grok Bot is packaging.
7. Cursor plugins ≠ Hermes skills hub. One is suite connectors; the other is portable procedures.
8. Always-on ≠ durable multi-role ownership. A VM that keeps clicking is not a work queue.

Why this is a good explanation

Definition: Overlay means placing each vendor feature in the layer whose job it performs, not the layer whose brand name it uses.

Hard-to-vary test: If you move Grok’s VM into “Model,” you cannot explain why work continues when the laptop closes but dies if the VM is killed. If you move Hermes Bot Mode into “Harness,” you cannot explain why turning the plugin off leaves profiles, cron, and CLI intact.[21]

Refutability: A future Grok API that exposes the loop, or a Hermes Bot Mode that owns a separate database, would force a remap.

Reach example: The same overlay is how you should read OpenClaw, Claude Cowork, Copilot agents, and AWS AgentCore: packaging vs loop vs host.

Criticism note: The house diagram still draws one Tools box labeled execution environment / safe sandbox. Grok Bot’s VM is too big for that box. Chapter 03 and 08 treat that as a stack miss, not a vendor miss.