The Agentic
Company

the allocation of permanence

The company does not become agentic by hiring a virtual employee for each box on an org chart. Its structure changes. Human judgment, trust, capital authority, accountability and relationships remain durable. Specialist capacity can be assembled for a bounded problem and dismissed when the problem changes. The output is useful now; the method, source trail and acceptance test should remain useful later.

That is a much harder design problem than “give every department a chatbot.” It forces a company to answer who is allowed to promise, what a delegated actor may change, which observations count as evidence, and where a human has to take the decision back.

five layers

The August 2026 operating-model report names founder authority, elastic capability, durable memory, commercial products, and external institutions. These layers are useful because they prevent a common category mistake: treating an agent’s completed task as a company’s accepted outcome. A processor, customer, regulator or partner can still disagree. A working artifact still needs an owner, a boundary and a route into a product or service someone values.

one working loop

A request arrives from a counterparty or the world. Capture the source before interpreting it. Compile it into the goal, constraints, identities, gaps and possible moves. Decide the few questions that carry consequence. Dispatch bounded work. Verify the outcome against the actual external system. Retain the method, asset and receipt so the next episode starts further ahead. The source report calls this the phone-native operating loop; the phone is a cockpit, not the company’s location.

the economics change with the risk

Affiliate distribution, services, infrastructure, performance participation and offer ownership are different economic positions. A company should choose among them according to which assets it owns, whose measurement it trusts, and what downside it carries. More automation does not make product liability, refunds, processor reserves or reputation disappear.

the deeper research

K-12 asks what makes one agent episode bounded and inspectable. G-9 asks how episodes compose. M-7 asks how agents from different organizations exchange value. The Institutional Kernel asks what the organization accepts as fact, commitment and authority. StandardKernel asks how observation becomes evidence, decision, execution, reconciliation and a receipt. These are connected questions, with different artifacts and different proof states. Follow each branch below rather than treating the synthesis as one already-deployed system.

where the org chart runs out

Take a customer problem that crosses marketing, support, payments and fulfillment. An org chart routes it through four departments. A loose collection of agents routes it through four chats. Neither structure tells you who owns the promise to the customer, whether a refund actually landed, or what the company learned. The company needs a place where those answers survive the people and agents who worked the case.

The division of labor starts with consequences. An agent can gather options, draft a reply, compare contracts or run a bounded test. A person who owns the relationship decides whether to make a new promise. An authorized finance actor decides whether money moves. A customer’s response, a processor record and a delivered package each answer a different part of the question, “did it happen?” Treat them as interchangeable and the company’s memory becomes fiction.

Human authority does not mean a founder reads every line an agent writes. That would turn the founder into the slowest API in the company. It means the company can name the few choices only that founder or another accountable person can make. Everything around those choices should be prepared so the human sees live options, evidence and consequence in one small view.

Elastic capability has a similar trap. A larger agent roster may produce more drafts and more coordination debt. The useful unit is a role with a bounded job: what it receives, what it may change, what would disprove its answer, what it must return, and who consumes the result. Another model or person can occupy that role when the roster changes.

follow one episode all the way through

Imagine the company hears that an order is missing. Capture the customer’s actual message and the order identifier. Compile the relevant state: carrier scan, payment record, fulfillment promise, refund rules, earlier correspondence. An agent might discover that the parcel is delayed and prepare three remedies. The decision owner picks one. Another actor sends the approved message and makes any authorized adjustment. Verification comes from the carrier, payment system and customer, as applicable. The retained method records how to handle this class of exception next time, including where the earlier process was wrong.

There is a serious difference between retaining the method and retaining the customer’s private details. A reusable procedure can describe the needed inputs, decision boundary and acceptance test without making another person’s data part of a public playbook. Institutional memory should make the next episode faster without widening access to the previous one.

Each layer exposes a different weakness. Strong models and weak memory mean the same expensive discovery gets made again next month. A beautiful knowledge base and weak authority mean an agent can act on stale permission. Strong internal coordination and weak external reconciliation mean the team celebrates work that a customer never received. The remedy depends on which layer failed.

choose the position, inherit the risk

An affiliate earns for distribution and depends on another party’s offer and accounting. A service provider owns execution and often inherits the support tail. Infrastructure earns from repeated use but takes on uptime, security and compatibility obligations. Performance participation adds measurement fights: whose baseline, which attribution window, which returns? Owning the offer captures more upside and more downside. These are business positions, not settings in an agent framework.

Better automation lowers the cost of making and testing a move. The economic question is where that capacity meets an owned asset, a credible promise and a buyer who will pay. The institutional question is whether the organization can remember what it promised and show what happened.

Pick one consequential episode from the past week. Can you identify the original signal, the person who owned the decision, the authority given to each actor, the outside record that closed the outcome, and the method that changed afterward? If any link requires interviewing whoever happened to be online that day, the company’s operating memory still lives in people’s heads. That is the part to build.