A company run by AI agents, steered by one human.
AI agents take over the classic business functions – product, tech, marketing, finance and legal. The human makes every call that matters. Agents only gain more autonomy once they have earned it (earned autonomy). Five apps in the stores show the efficiency gain of every iteration.
Most companies add AI to their org chart.
We drew the org chart for AI.
When work costs almost nothing to do, hands are no longer the scarce resource. Judgment, memory and trust are. So we organised around those three – and gave everything else to agents.
Agentic OrganizationA department is no longer a group of people. It's a mandate, a rhythm and a file.
The human doesn't do the work. The human owns the calls – and only those.
Autonomy isn't granted. It's earned, measured against human judgment.
A company that only learns from its mistakes ends up afraid of everything. Ours has to put a bolder move on the table every month.
Six principles.
Everything else is implementation.
A function is a file, not a person.
An agent's memory isn't a chat history – it's engineered. Every function keeps a file: mandate, rights, limits, open items, lessons. A run starts by reading it and ends by writing back. So any agent can take the chair – and nothing depends on who sat there yesterday.
Context EngineeringCan is not may.
Skills teach an agent how to do something. Whether it's allowed to is written in the mandate – and enforced by code, not by good intentions. An agent that knows how to upload to a store still can't.
GuardrailsHumans own the calls.
Three gates stay human: what gets built, what's good enough, what goes live. Everything in between runs on its own – often overnight.
Human above the LoopAutonomy is earned.
A step moves up only when the agent's judgment has matched the human's on the same cases, again and again. Some steps never move: taste isn't what gets automated, it's what the automation is for.
Earned AutonomyChallenge comes from outside the line.
A critic and an advocate review every function – and report to the human, not to the function they review. An agent supervising its own kind inherits its blind spots.
LLM as CriticLearn in three directions.
Every month: what to tighten, what to loosen, where to be bolder. All three need the human's yes. Otherwise a company only learns from its scars – and slowly stops daring anything.
Continuous ImprovementHow memory works
Not one endless chat history, but four layers – each loaded exactly when it's needed. That's context engineering.
The handbook
Maps, rules and lessons learned
every agent, at the startThe function's file
Mandate, open items, register
the agent in the chair, firstThe brief
Packages and checkboxes for one job – if a run dies, the next one picks up
the run working on itThe chronicle
Git history and board log: everything ever done
on demand, never wholeMemory you can read, version and audit – and that doesn't depend on who sat in the chair yesterday.
Five functions. Ten agents.
One human on top.
Every function is measured by the same yardstick: how far it has come from a first attempt to measured autonomy. None has reached the top yet. That's the work – and we show it as it is.
Click a function to see what's behind it.
The human
Priorities, the Go, money, the voice outside.
The human
Owns every call that matters – and only those.
- Priorities: the order on the board
- The Go: nothing goes live without it
- Start or stop a product
- Spending money
- Weekly brief
- Review pages for sign-off
- Exact go-live steps
- The Go
- Money
- Promises to customers
- Prioritises by drag and drop
- Signs off every story, image and voice
Fewer, better decisions – as agents earn autonomy.
Operating system
Runs the machine, keeps the human informed.
Operating system
Keeps the operating system running and the human informed. Summarises – never decides.
- Maintains board, control tower and maps
- Answers board comments
- Queues and starts runs
- Weekly brief
- Monthly learning loop: tighten, loosen, dare
- Model choice and token budget
- Sets priorities
- Changes a rule without the human's yes
feierabendarbeitsbeginnrechner-einrichten- Board with live gate and parallel runs
- Control tower: alarms, deadlines, store status
- Answers every board comment within a minute
The weekly brief and the learning loop on a fixed rhythm.
Independent challenge
Critic and advocate. Report to the CEO, not the line.
Independent challenge
Makes sure nothing important goes unchallenged – in either direction.
- Reviews specs, products and listings
- Writes blockers with evidence
- The bolder move every month (advocate)
- Stops or approves anything itself
- Reports to the function it reviews
lens: sceptical parentlens: sceptical buyer- First outing: 6 real blockers before a line of code
- Critic and advocate see the same facts
Measure the hit rate: how often a blocker was real.
Product
What gets built and why.
Product
Build the right thing.
- Research from store data and reviews
- Concept drafts in parallel strands
- Content production
- Sorts customer feedback
- Specs
- Pursue or kill an idea
- Design studies
- Decides scope
- Ships content without human sign-off
bilder-lokalpolly-audiocontent-nacharbeit- Every story passes an editor before image and voice
- 7,500+ voice clips, every one signed off
- Sign-off tools that never lose a verdict
A monthly review per app: what we assumed vs. what the numbers say.
Engineering
How it's built and shipped.
Engineering
Build it right and ship it safely.
- Cuts work into packages
- Builds, tests and commits
- Runs the guards
- Release bundles with the exact store steps
- Deploys a server
- Uploads to a store
- Publishes this website
nachtschichtaufs-phoneamazon-appstore- 121 commits in one night
- 4,300+ tests; QA caught 2 blockers 368 green tests missed
- Guards on production switches, memory pages, server version
A tech radar: platform deadlines, framework upgrades, crash rates.
Growth
How people find us – one brand, one voice.
Growth
The right people find the apps, and the brand speaks with one voice.
- Listing drafts and screenshots
- Website copy within the brand line
- Positioning
- Store experiments
- Posts in the founder's voice
- Speaks for the founder
- Spends money on ads
store-auftrittstore-texte- One store core feeds Google Play and Amazon
- A sceptical-buyer agent scores every headline
- Website checked on every release
Measure keyword ranks and store conversion every month. Paid ads later.
Finance
Where every euro comes from and goes.
Finance
Know where every euro comes from and where it goes.
- Collects receipts
- Watches deadlines
- Assigns costs to apps
- Monthly close
- Tax preparation
- Pays anything
- Files with authorities
- A missing invoice raises an alarm
- Deadlines in the control tower
- Revenue per store
The monthly close: receipts complete, revenue and cost per app, credits, deadlines.
Legal
No legal or policy risk goes unnoticed.
Legal
No legal or policy risk goes unnoticed.
- Scans primary sources
- Assesses what applies to which app
- Keeps the risk register
- Privacy updates
- Actions with a deadline, as board cards
- Decides a legal question
- Searches trademark registers by script
- Privacy policy and imprint for every app
- AI Act classification per app
A monthly regulatory radar: privacy, platform policies, consumer law – new, changed, emerging.
Agents are the staff, not the departments. A department is a file and a chair.
Same company.
Different physics.
Everything a classic company has still exists here. It just lives somewhere else.
Every app makes
the next one faster.
Dev days = days with commits, from the first commit to store submission. Store review time is not included.
How work flows
through the company.
Six phases. Every step is an agent with a clear role, a versioned skill and a definition of done. Agents run on their own – the human steers from above the loop, and steps into the loop at exactly three points: the concept, the sign-off and the Go.
Discovery
Research agents crunch store data, reviews and search demand. A critic and an advocate argue every idea – weak ones die by numbers.
Agentic WorkflowsagentsConcept
Two or three concept agents pitch in parallel strands. A critic tears the merged spec apart before a line of code exists.
Spec-driven Developmentfounder decidesBuild
An orchestrator cuts the work into packages and runs builder agents in parallel – often overnight, with a written brief and hard stop rules.
Multi-Agent Orchestration · Unattended RunsagentsCreate
Stories and scripts by Claude, checked by an editor agent. Illustrations from a local GPU pipeline, voice-over from Amazon Polly.
GenAI Content PipelineagentsQA
A QA agent checks the build against the spec, not just the tests. Every story, image and voice goes through human sign-off.
Human Sign-offhuman sign-offShip
Listings, copy and screenshots from one source of truth. Everything is prepared – then it waits.
Single Source of Truthwaits for GoAI does the work. Humans own the calls.
One board.
Every agent run starts here.
A self-built board is the cockpit. The founder prioritises by drag and drop; a card that is ready turns into an agent run with one click. Every move is logged.
A real card · Oct 9, 2026
A puzzle app, built overnight.
Replayed from the git log of our puzzle book app. Shortened and translated, times are real.
Not everything should run alone.
A step moves up only when the agent's output has matched human judgment on the same inputs, again and again. Some steps are meant to stay at “AI drafts, human decides” – judgment is not what gets automated, it's what the automation is for.
| Step | Human does it | AI assists | AI drafts, human decides | AI runs, human above the loop | Why |
|---|---|---|---|---|---|
| Discovery | Kill or pursue is a founder call | ||||
| Concept | Agents pitch and criticise, the founder picks | ||||
| Build | Specs, tests and guards make it safe to run alone | ||||
| Create | Every story, image and voice is signed off | ||||
| QA – tests | Thousands of tests and a spec check per build | ||||
| QA – taste | “Good enough for a child” stays human | ||||
| Ship | Prepared by agents, released on a Go |
Agents are fast.
The guardrails are the method.
Nothing ships by itself
Runs prepare everything and stop at “Go live?” with the exact steps written out. Release is a human decision, every time.
A card is a change, not a release
Changes collect on main. Builds, version numbers and store uploads only happen on a release button.
Watchers on the shipping state
Tests that guard what goes to the store: production switches, memory page sizes, server version. Fail one and the bundle is never built.
Sign-off that can't get lost
Human judgment is the bottleneck, so every review click is saved the moment it happens.
State lives in files
Every unattended job has a brief with packages and checkboxes. If a run dies, the next one picks up where it stopped.
Context is engineered
Maps, rules and lessons live in versioned files every agent reads – not in someone's head, not in a chat history.
What runs the studio.
Build
- Claude Code
- Claude Opus
- Claude Sonnet
- Versioned skills
- Flutter
- GitHub
Inside the apps
- Claude on Amazon Bedrock
- AWS Lambda
- Amazon DynamoDB
- On-device speech recognition
Create
- Local NVIDIA RTX
- Open image models
- Amazon Polly
Operate
- Self-built agent board
- Sign-off tools
- Watchers
- Control tower
Ship
- Google Play
- Amazon Appstore
Patrick Steinborn
Product leader with more than 20 years in consumer internet, most of them at VP level – growth, monetisation, pricing and AI transformation for products with millions of users. He founded Remindery in 2026 to prove a simple thesis: with the right operating model, one person can run what used to take a whole team. He builds every day, with Claude Code as his main tool.
Curious how this could work for your company? Let's talk.