Company OS Learnings — What I Kept After Retiring the Control Plane
Status (Sep 2026): Company OS is a historical thesis, not the current ops platform. The learnings below come only from what is already on this site — the six-layer model, system map, Coolify heritage, and operating notes — not from new metrics or internal dashboards.
When I named Company OS, I was trying to make coordination programmable: priorities, context, agents, tools, guardrails, and feedback in one loop. TUR Company OS wired real providers — Supabase, Linear, Inngest, PostHog, Raindrop, and others — behind adapters with explicit trust boundaries (Operating Note 001).
That experiment taught me plenty. It also grew heavier than a solo portfolio needed for shipping this site. Day-to-day work now runs through Cursor Cloud Automations on this site and the stack on Toolstack. The thesis page stays live for SEO and reference; the narrative is superseded, not deleted.
What changed
| Before | Now |
|---|---|
| Company OS as current ops framing in nav and profile | Toolstack in top nav; profile currentFocus points at the live stack |
| TUR control plane as the default mental model for “what runs” | Issue → label → Cloud Agent → draft PR → CI → human merge |
| Self-hosted Coolify / n8n / LiteLLM / Paperclip as execution heritage | Still documented on /company-os and in older stack posts — not the primary loop for this repo |
For how Company OS differs from agent frameworks and orchestration tools, see Company OS vs agent framework vs orchestration. For an earlier self-hosted 24/7 stack (LangGraph, Paperclip, LiteLLM on Hetzner), see AI agent stack as competitive edge — a different era, not a verdict on either approach.
Seven learnings I am keeping
1. A model is not an organisation
The /company-os premise still holds: a model is a capability, an agent is a worker, neither is an organisation. Useful systems need explicit intent, ownership, and feedback — not another chat window.
2. Six layers beat “one AI agent”
Intent, context, agents/workflows, tools, approvals, and outcomes each fail differently. Collapsing them hides the interfaces you must design, secure, and measure (six-layer reference).
3. Adapters beat permanent dependencies
The live system map treats providers as replaceable: Company OS owns state, policy, and memory; Supabase, Inngest, and the rest stay behind adapters. When the control plane moved to Cursor, that lesson transferred — swap the runtime, keep the boundaries.
4. Autonomy is a ladder, not a switch
Assist → Execute → Coordinate → Operate → Adapt. High autonomy belongs in well-observed, reversible work; money, publishing, and production changes keep human gates (autonomy ladder).
5. Coolify heritage is real execution memory
The self-hosted stack (Coolify, n8n, LiteLLM, Paperclip) remains part of the execution story on the thesis page. It proved durable workflows and routing on owned infra; it did not have to remain the default loop for every repo.
6. Operating notes beat product theatre
Operating Note 001 and the comparison post aged better than pretending the map was finished. Publishing what is wired and what is staged beats announcing a product that is still a thesis.
7. Smaller loops can ship the narrative
Retiring Company OS from top nav does not erase the architecture. It narrows what “current” means: Toolstack for tools, Cursor Automations for this site’s change loop, /company-os for the archive.
Where to go next
- Architecture archive: Company OS — system map, six layers, autonomy ladder
- Active loop: Cursor Cloud Automations on this site
- Tools in use: Toolstack
- Broader agentic framing: What is agentic engineering?
For the systems behind this work, explore the projects, the toolstack, or start a conversation.