Teams are dead.
The factory is the
operating system.
A framework for the transition from individual AI productivity to Institutional AI — where agents, humans, and guardrails form a single operating system.
Read the thesis →Everyone is using AI to write code faster. Almost nobody is asking the harder question: what happens when the organisation itself becomes the machine?
The Core Thesis
Most companies are stuck giving developers better autocomplete and calling it transformation. They are optimising the typist. The unit of productivity is still the individual.
The Factory Way is a different bet. What if high-quality software is a systems problem, not a people-and-talent problem? What if the right unit of productivity is the organisation, not the developer?
We call this Institutional AI — where agents don't assist individuals, they operate as an organisational system. Humans govern, direct, and review. The factory ships.
The goal is not efficiency. The goal is to compress the time from idea to live, at a quality level that was previously impossible without an army.
Individual AI vs Institutional AI
Where most teams are
- +98% More PRs opened
- +91% Longer review times
- +47% More context switching
- Flat Delivery metrics unchanged
- Each developer has their own AI setup. Agents don't know about each other. The pipeline clogs.
Where the future lies
- 1,300 PRs/week, zero human code (Stripe)
- 50%+ Of merged PRs from agents (Ramp)
- 70% Machine-generated code (Uber)
- 21K Developer hours saved (Uber)
- Agents operate within institutional infrastructure. Humans govern, direct, and review. The factory ships.
The 12 Levels
Individual AI
A developer uses AI to amplify their own output. Autocomplete, chat assistants, agentic coding tools. The human drives everything.
Transitional
The organisation builds shared AI infrastructure. Work remains human-driven, but the tooling is institutional. The bridge between two worlds.
Institutional AI
Agents operate as an organisational system. Humans govern, direct, and review. The factory runs. This is where the future lies.
Surprising Findings
Experienced developers were slower with AI tools in METR's controlled trial. They believed they were 20% faster — a 39-point perception gap.
Read more on L4 →Teams at L8 produce 98% more pull requests — but delivery metrics stay flat. Review time up 91%. Context switching up 47%. This is the trap.
Read more on L8 →Stripe's Minions ship 1,300+ PRs per week with zero human-written code. Built on a decade of standardised devboxes and 3 million tests.
Read more on L10 →Nearly half of AI-generated code contains security vulnerabilities. Self-healing loops make it worse — 37.6% more critical vulnerabilities after 5 iterations.
Read the research →Where is your organisation?
Answer 5 questions to find your approximate level on the framework.
How does your team primarily use AI in development?
Non-Negotiables
Without guardrails, AI work is a liability.
In regulated industries, unguarded agent output isn't innovation — it's risk. Guardrails are a first-class organisational function, not a compliance afterthought.
Evidence over assertion.
Verification requires tool logs and linked artifacts — not agent self-reporting. If it can't be traced, it didn't happen.
Stay in the cheapest reasoning space.
Planning is cheap. Execution is expensive. Most teams burn resources executing a mediocre plan when they should still be iterating on the idea. Stay in plan space until the thinking converges.
The factory learns.
Knowledge lives in the improvement cycles. Every task deposits insight. The reflect cycle turns raw experience into reusable intelligence. Prompt study — understanding what works and why — is the discipline that compounds.
Intent persistence is a first-class concern.
"The human asked me" is not intent. The system captures WHY, not just WHAT.
The factory thinks in capabilities, not brand names.
No single model failure should break the system. The factory references capability tiers and routes accordingly. Your intelligence layer should survive any vendor's bad quarter.
Time from customer request to deployed capability in a new market.
Building for enterprises where
failure has real consequences.
The path from wherever you are on the 12 levels to Institutional AI requires more than better tools. If this resonates, we want to hear from you.
Get on the list
The transition to Institutional AI is coming. Leave your email and tell us what you're most interested in.
Let's talk.
Whether you're exploring AI adoption, stuck at L8, or ready to build institutional infrastructure — we want to hear from you.