HQ for Work

Why Shadstone should review this

We’re still deciding how a distributed online team shares AI agents — not five people each reinventing strategy in private Claude/Codex threads, and not a single “god chat” nobody trusts. HQ is explicitly aimed at that gap: one coordinated system under the tools people already use.

This page is a team evaluation brief, not a purchase decision. Read it next to our existing map: multi-agent harness, multiplayer agents, QM, company brain.

Source

Jacob Posel (@jacob_posel) — building @HQForWork, “the shared AI context layer for your business” (hqforwork.com).

Tweet that sparked this page: x.com/i/status/2086844678486274106 (10 Aug 2026) — “This has transformed our company,” quoting a long-form post by @VibeMarketer_ on turning scattered models/agents/skills into one coordinated system with HQ.

Transparency: Jacob is the founder of HQ; some readers will treat the thread as product marketing. The underlying harness pattern (memory → policy → worker → review → sync) is still worth evaluating whether you adopt HQ, QM, HEOS-style vaults, or a hybrid.

One-sentence TL;DR

Most teams have smart models trapped in private, diverging contexts; HQ claims to sit underneath Claude Code / Codex / Cursor (or open-source models) as a shared filesystem-shaped company layer — knowledge, policies, workers, and a review→sync loop so one person’s improvement becomes the team’s next starting point.

The problem (exactly Shadstone’s pain)

One person briefs Claude with the latest strategy. Another asks Codex to search an old folder. A third rebuilds a useful workflow from memory. Every chat contains a slightly different version of the business.

That is distributed-team default mode:

• Context lives in individuals’ chats, not the company
• Corrections don’t propagate (you re-teach the same judgment weekly)
• Skills and automations never become team infrastructure
• Model upgrades reset “how we work” because the company never lived outside the prompt

HQ’s answer: the model supplies intelligence; the harness supplies the company.

What HQ is (in plain English)

Install path mentioned in the long post: guided setup or open-source quick start npx create-hq — confirm current docs on hqforwork.com before a pilot.

The five questions a company harness must answer

From the VibeMarketer / HQ write-up — useful even if we don’t pick HQ:

1. What does the AI know? (company memory)
2. How does it find relevant context? (routing / charter / search)
3. Which rules must it follow? (policy)
4. What repeatable work can it perform? (workers)
5. How does each run improve the next one? (review → sync)

A long system prompt answers some of these for one session. A company harness makes them structured, persistent, and available to everyone.

Prove it on one workflow (don’t map the whole company)

Explicit advice: do not begin by modeling all of Shadstone. Pick a workflow that (1) happens often, (2) has clear boundaries, (3) needs company context, (4) a human can judge quickly.

Their example: weekly company-intelligence brief — meetings + projects + decisions + risks → sourced draft → human review before distribution.

For Shadstone, first candidates might be: weekly SBU/partner brief, CBS pipeline digest, Para Living / theflysales product pulse, or “what shipped on mikesblogdesign this week” — pick one, make it boring and testable.

Memory → policy → worker → correction layers

Memory   company-brief, projects, decisions, people, meeting notes
   ↓
Policy   how work must be done (cite sources, no secrets, approval gates)
   ↓
Worker   narrow reusable procedure (/newworker)
   ↓
Run      draft for human review
   ↓
Diagnose which layer failed (knowledge vs routing vs skill vs policy)
   ↓
Sync     reviewed improvements → Main → team default

Correction routing (this is the compounding bit):

• Missing fact → knowledge path
• Wrong context → routing / resource descriptions
• Repeated mistake → worker skill
• Unsafe behavior → policy or mechanical hook
• Weak deliverable → output contract
• Stale info → knowledge gardening

Three levels of control: instruction (preference) → policy (durable) → hook (blocks costly actions). “Cite sources” can be policy; “never send without approval” needs an action gate.

Do we already have pages on this? Yes — here’s the map

Shadstone has already been thinking in harness terms. HQ is a product candidate that sits in the same category as several pages we already maintain:

Chat surface     Slack · Buzz · GFAVIP · GrokBot
Harness runtime  QM · OpenClaw · Hermes · Grok Build
Shared context   HQ · HEOS/vault · Company Brain files   ← this page
Tools / compute  CLIs · MCP · exe.dev VMs

HQ vs alternatives (for Shadstone decision)

Shadstone evaluation checklist

For the next team review (async Slack or live):

1. Tenant boundaries — Can Para Living, GFAVIP, client work stay isolated? Shared harness only works when “shared” is explicit (the post stresses company isolation).
2. Who may sync to Main? — Review before merge; treat like production deploys.
3. First worker — Name one weekly brief; write inputs/output/approval boundary before install.
4. Tool fit — Do ninjas live in Claude Code, Codex, Cursor, Grok Build, or mix? HQ’s pitch is model-agnostic underneath — verify with our actual stack.
5. Secrets — Secrets manager claims exist in HQ’s product story; map to our current secret handling.
6. Compare to QM + vault — Same pilot workflow in HQ vs git-based vault + Grok Build; score after 2 weeks.
7. Human gates — Distribution, spend, client comms stay draft-only until named humans approve.

Suggested pilot (2 weeks)

1. Install HQ (or clone equivalent vault structure) for one Shadstone workspace only.
2. Seed: short company-brief.md + 3–5 project files + meeting notes path for one week.
3. Policy: weekly-intelligence.md (cite sources, flag uncertainty, no cross-tenant, no send).
4. Worker: weekly-intelligence; run Monday; founder/ops reviews; fix layers not prompts.
5. Second human runs the same worker Friday without private coaching — if they fail, harness isn’t ready.
6. Write a one-pager: keep HQ / try QM / stay vault-only / hybrid — then expand to a second workflow (e.g. product brain for theflysales).

Progression (quote-worthy)

Memory → context → policy → worker → review → team default

Start with one recurring job this week. The model will keep changing; the company’s memory, rules, and best ways of working should not reset with it.

Related on this site

• Multi-agent harness map — category hub (add HQ here mentally as shared context)
• Multiplayer agents — many humans, one shared agent memory
• QM — YC open-source multiplayer harness alternative
• Company brain — org-scale single brain narrative
• Harness tips · Agentic setup
• Buzz · Open Tag · Claude Tag
• GrokBot · God agent vs team
• Product brain — per-SKU memory (can be a worker family under HQ)

Primary sources: Jacob @jacob_posel · @VibeMarketer_ long post · hqforwork.com