The AI company model
Black Box is an AI company, not an AI assistant. That one distinction drives every product decision — from the UI to the price, from how it plans work to how it recovers from errors.
TL;DR
An AI company is a software product that behaves like an organization: specialists with distinct roles, a chain of command, memory that persists across sessions, and the authority to take real actions on behalf of a business. Three marks separate it from everything nearby — specialization, delegation, and authority.
The working definition
An AI company is a software product that behaves like an organization. Three marks separate it from an AI assistant, a lone agent, or a workflow tool:
- Specialization. Not one general model trying to do every job. Instead, a CEO who coordinates specialists — each with a role, a tool belt, and their own memory.
- Delegation. The owner hands over outcomes, not tasks. "Ship the newsletter" — not "write an email, then copy it, then paste it into Mailchimp."
- Authority to act. The team takes real actions — writes files, calls APIs, sends emails, deploys pages — not just text responses.
An AI assistant is one model behind a chat box. You are the orchestrator: you ask, it answers, you decide what to ask next, you paste the output somewhere useful. An AI company flips the contract. You name the outcome; the team figures out the moves.
Why the organizational frame
We could have called Black Box a "multi-agent system," which is accurate but useless. Multi-agent systems are a research primitive. They do not have titles, they do not ship work, and they have no place for a human owner to sit.
The organizational frame is what lets a solo consultant — someone with no engineering background, no prompt-engineering practice, and no patience for yet another SaaS with wiring to do — actually get work out of a multi-agent system. She knows what a CEO is. She knows what a head of marketing does. She does not need to learn any new vocabulary, and she does not need to know the word "agent."
The organizational frame is also what makes the pricing make sense. You are not paying for tokens. You are paying for a team. A $500/mo team that replaces $10K of human headcount is obvious value; a $500/mo subscription to "an AI product" is not.
What changes when you take the frame seriously
If you commit to the AI company frame, product decisions stop being hard. Everything follows.
- UI becomes Boardroom + inbox, not a chat box. You do not talk to your whole company at once. You read reports, approve decisions, and have one-on-ones. See the four views.
- Each specialist gets its own memory and tool belt. The Coding Specialist knows git and Railway; the Legal Specialist flags everything for human review. Forcing one prompt to do both regresses both.
- The CEO is a real abstraction, not a routing function. The CEO decomposes goals, delegates, waits for results, and reports back. It surfaces you only for decisions only you can make.
- Errors are organizational, not technical. "Coding Specialist hit a rate limit, re-queued to the next window" is a real report. "Tool call 7 failed, retry" is not.
How it works in Black Box
Black Box ships the AI company frame in code. The CEO loop runs
on @anthropic-ai/claude-agent-sdk
as a single long-running query() (see
apps/engine/src/ceo/loop.ts). Each specialist is
delegated as a separate query() sub-session with a
scoped system prompt, its own allowed-tool list, its own MCP
servers, and its own working directory. The CEO is literally a
different process from the specialists it spawns; the
organizational structure is enforced at the process boundary.
An Evaluator sub-agent reviews each deliverable before it reaches you. A non-LLM circuit breaker watches every task and halts on five safety conditions. All of that maps directly to the frame: a company has QA, a company has a fuse box.
What it is not
Black Box is not a replacement for a human team of fifteen people. A senior engineer with three years of context on your codebase will still outrun a Coding Specialist on a non-standard refactor. What Black Box replaces is the marginal hire: the VA you considered, the freelancer you keep re-briefing, the junior dev you do not want to manage.
It is also not a general-purpose AI. It refuses to roleplay, it declines off-scope tasks, and it stays narrow on purpose. Wide is someone else's category.
Where the category is going
The first wave of AI companies is vertical — an AI company for consultants, one for real-estate agents, one for restaurants. The second wave will be horizontal: the same orchestration, the same specialists, with skill packs for each vertical. Black Box is being built for the horizontal future while shipping the consultant wedge today.
Frequently asked
- Isn't an AI company just agents with a coordinator?
- Technically yes. But the organizational framing is what lets a non-technical owner use it. That framing is the product.
- Is "AI company" just marketing?
- No — it is a product category with three load-bearing traits: specialization, delegation, authority. Miss one and you have an assistant, an agent, or a workflow.
- Is this the same as agent swarms?
- No. Swarms are flat and emergent. An AI company is hierarchical by design — a CEO, specialists, and an owner.
Related concepts
- The CEO agent — the spine of the company.
- The 18 specialists — who reports to the CEO.
- Delegation — how work moves from CEO to specialist.
- Essay: what is an AI company? — the long form.
- Glossary: AI executive team.
Next: meet the CEO
The CEO agent is the one agent you actually talk to. Here is how it decides.