How the factory runs a company
People decide where the company goes. The factory does the work in every function, checks it, and gets better at it on its own. Leadership sees all of it from one place and is asked only about what is theirs to decide.
Seven layers
Each layer has one job. Mixing them is how automation turns into a pile of scripts nobody trusts: the agent that does the work never also decides whether it is done, and nothing below the top layer sets direction.
- 1DirectionPeople
Leadership and a few key people set goals, limits and who may decide what. Nothing else in the factory sets direction.
- 2Control towerEveryone can look
One view of the whole company: outcomes, queues, wait times, spend, risks, and the short list of decisions that need a person.
- 3Shared contextFactory
The state of every piece of work, the knowledge it needs and the current version of every procedure. Agents read from it and write back to it, so nothing lives in someone’s head or inbox.
- 4OrchestrationFactory
Breaks goals into jobs, orders them by dependency, caps work in progress and keeps every job inside its budget and permissions.
- 5ExecutionAgents, by function
- SalesQualifies leads, drafts proposals, follows up after calls
- MarketingWrites and tests pages, campaigns and articles
- Product and engineeringSpecs, code, tests, releases
- SupportAnswers, reproduces bugs, files them with evidence
- FinanceInvoices, reconciles, forecasts spend
- OperationsOnboarding, access, vendors, internal tools
- 6VerificationIndependent checks
Every result is checked by something that did not produce it: tests, a second model, a rule, a reconciliation. Claims count for nothing; evidence decides.
- 7LearningFactory, approved by owners
Every failure and every person’s correction becomes a proposed change to a procedure. It is tested on past work, rolled out, and rolled back if the numbers get worse.
The view from the control tower
Leadership does not read reports about the company; it looks at the company. Every job, queue and dollar is in one view, and the screen ends with the decisions waiting for a person. Most days that list is short.
- Jobs in flight214across six functions
- Waiting for a person6oldest 3 h
- Median time to done3.4 hlast 7 days
- Spend today$1,840of a $2,500 cap
Work by function, today
- Sales1 waiting
- Marketingclear
- Product and engineering3 waiting
- Support1 waiting
- Finance1 waiting
- Operationsclear
Decisions waiting for a person 3 of 6
- Approve a 15% discount for a three-year contractAbove the sales limit of 10% · CEO
- Ship the new onboarding to all customersVerification passed; rollout is a product call · Head of product
- Raise the support budget for this weekTicket volume is 2× the forecast · COO
What the factory learned this week
- Proposal procedure v14: adds the security questionnaire up front. Tested on 40 past deals, rolling out.
- Release checklist v31 rolled back: slower reviews, no fewer defects.
What runs through it
Every function hands the factory the work that follows a procedure and keeps the work that needs judgement, relationships or the authority to commit the company.
| Function | The factory does | A person does | The wait that disappears |
|---|---|---|---|
| Sales | Research on every lead, call notes, proposals from approved templates, follow-ups, CRM hygiene | Price exceptions, relationships, the final yes | A proposal waiting a week for someone to write it |
| Marketing | Pages, articles, ad variants, A/B tests, weekly reports on what worked | Positioning, brand, what the company refuses to say | Campaign ideas queued behind one designer and one copywriter |
| Product and engineering | Specs from tickets, code, tests, review by a second model, releases behind flags | What to build next, product taste, risky migrations | Pull requests waiting days for a reviewer |
| Support | Answers from the knowledge base, reproducing bugs, filing them with logs and steps | Angry customers, refunds over a limit, policy changes | Bugs bounced between support and engineering for lack of detail |
| Finance | Invoices, reconciliation, spend tracking per job, monthly close drafts | Budgets, payments above a limit, anything that signs for the company | Month-end close as a two-week project |
| Operations | Onboarding checklists, access requests within policy, vendor renewals, internal tools | Hiring decisions, contracts, exceptions to policy | New people waiting days for accounts and access |
One job, end to end
A request that used to need five meetings crosses five functions in a day. Nobody forwards anything: each step reads the shared record, does its part and writes the result back, and the one human decision happens in the control tower.
One request, five functions, no hand-offs by email
- Agent
- Person
- Automated check
Steps as text
| Step | What happens |
|---|---|
| 1. A customer asks for single sign-on on a sales call | Agent — Sales agent writes the request into the shared context, with the call excerpt |
| 2. Product drafts the spec | Agent — Acceptance criteria, effort, and which plans get it |
| 3. Head of product says yes | Person — One decision, made from the control tower in two minutes |
| 4. Engineering builds it behind a flag | Agent — Code, tests, docs; a second model reviews the diff |
| 5. Verification | Automated check — Tests pass, a scripted login works against a test identity provider |
| 6. Support, finance and sales follow | Agent — Help article updated, plan price changed, the customer gets a reply with a date |
Failure is part of the design, not an exception to it. Every job has a retry budget, a money budget and a clear point at which it stops and asks.
When the job fails
- Automated check
- Agent
- Person
Steps as text
| Step | What happens |
|---|---|
| 1. Verification fails | Automated check — the scripted login times out |
| 2. Agent retries | Agent — with the failure in context |
| 3. Second failure | Automated check — same check |
| 4. Job parks | Agent — state saved, budget frozen |
| 5. Owner gets one question | Person — the identity provider rejects our test tenant: which account? |
| Loop | one retry with the evidence: back from step 2 to step 1 |
Where work used to wait
Fast agents expose every slow step around them. The factory is mostly a set of answers to the places where work waits for a person.
Hand-offs between teams
One shared state for every job. The next function picks the work up from the record, with its history, instead of from a message.
Approvals
Delegated limits. Owners set the bounds once (a discount, a budget, a class of change); inside them the factory acts, outside them it asks.
Review of everything
Independent verification by default and human review by risk. People read the changes that can hurt, not every change.
Too much work started at once
Work-in-progress limits per function. New work waits in a queue the control tower can see, instead of half-finishing in ten places.
Knowledge in people’s heads
Procedures and decisions written into the shared context, because agents can only follow what is written. The side effect is a company that can explain itself.
The same mistake twice
Every correction becomes a tested change to a procedure, so the next job starts from the fixed version.
How it learns
Learning is not a model getting smarter by itself. It is procedures that change only when a change is shown to work, with a version number and a way back.
How a procedure gets better
- Automated check
- Agent
Steps as text
| Step | What happens |
|---|---|
| 1. Outcome | Automated check — a failure, a slow job, a correction |
| 2. Diagnosis | Agent — which step and why |
| 3. New version | Agent — a proposed procedure change |
| 4. Replay | Automated check — run on past jobs |
| 5. Rollout | Agent — a share first, then all |
| Loop | rolled back if the numbers get worse: back from step 5 to step 3 |
What stays with people
Direction
Which markets, which products, what the company will not do.
Authority
The limits inside which the factory may act, and who may change them.
Taste
Whether the work is good, not only whether it passes the checks.
Trust
Customers, partners and teams still deal with people they know.
The parts I build
The engine at the centre of my work is Bulava, where Claude Code builds and Codex reviews until the work is accepted. One measured example of a single part: on the same game-rebuild task, two independent model positions before any code took 1.5 hours and 34.3 M tokens against 3.7 hours and 389 M for Claude Code alone. It is one coding task, not a whole company, and the method is public. I’m also building a dark factory at Raccoon Gang.
