Just when we figured out how to prompt agents to do our work for us, that approach is already obsolete. I spent the last few months proud of running ten parallel Claude sessions. Turns out I was already a generation behind.
People like Boris Cherny, Peter Steinberger, and the wider San Francisco crowd say they are not prompting Claude anymore. They build loops and workflows - and let agents prompt themselves, evaluate themselves, and optimize themselves. Y Combinator has a name for this: the closed-loop organization.
The gap is not about tools. It is about architecture.
What a closed loop is
In a closed loop everything is recorded and everything is in agent-readable form. Agents evaluate their own outputs against objectives. When they miss, they self-correct. It is reinforcement learning - applied not to a model, but to an organization. Y Combinator is already backing startups built entirely around this pattern.
An open loop is a company where a human assigns a task, an agent does it, and a human checks the result and hands it off by hand. It is still a human orchestra with AI tools inside. It works, but the ceiling is set by how many loops one person can run - my ten parallel sessions, for example. The closed loop removes that ceiling: the system does the checking and the correcting itself.
Reinforcement learning applied not to a model, but to an organization. A machine that measures itself, corrects itself, and improves itself.
What this does to org design
Humans do not disappear. They move to the edges. They become the front end of the company - the interface with customers and with the outside world. What gets erased are the middle layers:
- The human middleware ends. Managers who existed mainly to coordinate, translate, and relay information between levels lose their reason to exist. The loop does that work.
- Two kinds of roles remain. ICs who can direct agents, and DRPs - directly responsible persons - who own the outcome and let the system handle execution.
- The cost structure shifts. Less payroll, more tokens. The cost that used to sit in headcount moves into compute spend on the P&L.
The uncomfortable question for large enterprises
And here is the uncomfortable question for those of us in large enterprises: we are still debating whether to let employees use AI tools at all. The next generation of companies will not have this conversation. They will have systems that prompt themselves, humans at the edges, and token spend on the P&L where headcount used to be.
The gap between an open loop and a closed loop does not close by buying another tool. It is a difference in the architecture of the company. And that is not something you catch up on overnight.
If you are thinking about how to move your organization from an open loop closer to a closed one, get in touch. I am happy to look at where your loop still ends at a human and where a system could close it.
