5 min read Rocky Elsalaymeh
Run an AI company. Not a prompt.
Prompt engineering is a dead end. The interesting unit of work is no longer the prompt; it is the org. Here is the case for treating AI like a company, and the architecture that makes it real.
Two years into the agent gold rush, the dominant interface for working with AI is still a chat box. You write a prompt, you read a wall of text, you copy a paragraph into a doc, you start over tomorrow. The prompt is a unit of memory the size of a Post-it note, and the entire industry has spent three years optimizing the handwriting on it.
That is the dead end this product walks away from.
What the chat box hides
The chat box is a great surface for retrieval. It is a terrible surface for operations. Operations have shape: roles, hierarchies, queues, goals, deadlines, deliverables, post-mortems. None of that fits inside a turn-by-turn dialogue.
The dominant frameworks have tried to bolt operations on top of the prompt. LangChain put a graph around it. AutoGen put a roundtable around it. CrewAI put a script around it. Each of them is a thin layer of orchestration on top of a substrate that was never built to host an organization. You end up with a Python file that pretends to be a company.
Real organizations look nothing like a Python file. They look like an org chart, a kanban board, a meeting calendar, and a vault of artifacts, all running concurrently against a shared body of context. That is the substrate this product builds.
The unit of work is the role, not the prompt
The most important architectural decision in Team-X is that the primitive is a role, not a prompt. There are 57 of them, hand-written. Officer, Senior Management, Management, Supervisor, Lead, IC: six levels with real authority gradients. Each role spec defines the agent’s lens: what they care about, what they can do, what they delegate, what they refuse.
Roles compose into an org. The org chart is editable. You can promote a senior engineer into a director, fire a contractor, hire a copywriter from the catalog, redraw the reporting lines. The state of that org persists in SQLite, not in a prompt. Every employee carries their full history through every conversation they have, not just the last 8000 tokens.
The benefit shows up the moment you ask a non-trivial question. “Why is the frontend team behind schedule?” The chat-box answer is a hallucinated narrative. The org-substrate answer reads the actual ticket queue, the actual project deadlines, the actual cost telemetry, and the actual meeting minutes for the last two weeks, then writes the post-mortem citing specific tickets and people by name. The first system is doing creative writing. The second is doing analysis.
The cockpit, not the conversation
Once you accept that the unit of work is a role and the substrate is an org, the chat box stops being the right shape for the interface. What you want is a cockpit: an environment where you can see the whole company at once, intervene anywhere, and watch the work happen.
That is the surface Team-X ships. Five subviews on the dashboard. A live token stream from every employee. A four-column kanban with automatic agent assignment. A real schedule with due dates, project targets, goal targets, and overdue counters. A meeting room you can pull anyone into with one click. A natural-language command palette: hire someone, file a ticket, call an all-hands, ask a question, all from Cmd+K.
The interesting design constraint is that none of this is a feature on top of a chat. It is the chat that is a feature on top of the org. The Chat tab exists; it is one of nine. The center of gravity has moved.
You don’t manage prompts. You run a company.
Local-first is a posture, not a feature
The other architectural decision worth defending is the privacy posture. Team-X runs fully on-device. Ollama integration is first-class, not an afterthought. The recommended setup is llama3.1:8b running locally, with the optional ability to add cloud providers if and when you want to. There are 10 supported providers across three tiers: Local, Open-Source Cloud, Proprietary Cloud. You filter which tiers your agents are allowed to use.
There is no analytics. There is no telemetry. There is no auto-update phone-home. Updates are user-triggered, pulled from GitHub Releases on demand. Secrets live in the OS keychain, not in plaintext config files. The only way the app talks to the network is the way you tell it to.
This matters because the obvious alternative, the one most of the industry has chosen, is to ship a hosted SaaS that ingests your operational data into someone else’s data warehouse. Once your tickets, meeting minutes, and post-mortems live there, you are not running an AI company; you are renting one, and the lock-in is permanent. Cloudflare’s 2024 analysis of agent-platform terms of service made the dynamic explicit: vendor SaaS captures the operational substrate, then charges rent on it forever.
Local-first inverts that contract. Your data is yours. Your roles are yours. Your org chart is yours. The model providers are interchangeable. The whole thing is MIT-licensed.
What ships on day one
Team-X v3.0 ships:
- The full org with 57 hand-written roles
- The five-subview live cockpit
- The four-column kanban with auto-assignment
- One-click meetings with auto-generated minutes and action item extraction
- The natural-language command palette with 14 structured intents and an agentic-loop fallback
- A ReAct-style agentic loop dispatching six read-only query tools against the org
- A copilot service that asks “what’s wrong with this company right now?” on a five-minute cadence and surfaces deduped insights
- A file vault with SHA256 integrity verification and FTS5 search
- Cross-platform installers for Windows, macOS, and Linux
- 2,150 passing tests on commit zero
The repo is open. The license is MIT. The first commit is the launch.
What this is not
It is not an autonomous-agent demo. It is not a chat-app skin. It is not a wrapper that calls GPT-4 and asks it to roleplay as a company. The agents are real, the org state is real, the deliverables are real files in your vault, and the cost reports are real numbers from your provider invoices.
The thesis is simple: if AI is going to be useful for operations, the interface has to look like operations. That means an org, not a prompt. A cockpit, not a chat. A company, not a conversation.
The repo is at github.com/git-rocky-stack/team-x. The download is on /download. The role catalog is browsable at /roles. The privacy posture lives at /privacy.
If you want a weekly note from the shop floor as Team-X grows, subscribe to Field Notes below. Every Thursday. No tracking. RSS available.
You don't manage prompts. You run a company.
A note from the shop floor. Every Thursday.
Check your inbox. One-click confirm link on the way.
Something went wrong. Please try again, or email notes@team-x.app directly.