Organizations are built from materials.

Not only physical materials like offices, warehouses, and servers, but materials for action. People who can interpret a situation. Rules that constrain what they may do. Software that repeats a procedure. Records that preserve what happened. Roles that determine who has authority. Reviews that turn one person’s judgment into a decision the institution is prepared to stand behind.

For most of modern organizational history, these materials fell into a rough division. Humans provided flexible reasoning. They interpreted ambiguous circumstances, made tradeoffs, noticed exceptions, and acted when the procedure ran out. Rules, machines, and later software provided repeatability. They stabilized what could be specified in advance and executed it with greater consistency than a person could.

The division was never clean. People follow routines. Software contains judgment frozen into code. A form can create discretion while a meeting can become mechanical. But the distinction was stable enough to organize around. We knew where interpretation was expected to happen and where deterministic execution was supposed to begin.

AI agents introduce a different material into this arrangement. They can interpret instructions, navigate ambiguity, use tools, adapt a plan, and produce work that looks like the output of situated judgment. But they do not arrive with the institutional properties that historically surrounded a human reasoner. They do not have professional standing, personal liability, organizational loyalty, a career to protect, or a place in a chain of responsibility simply because they can reason.

We have made discretion computational before making accountability computational.

That is the design problem underneath agentic organizations. It does not live entirely inside the model, the human, or the software. It lives in the structure between them.

The Old Materials of Organization

Consider how a conventional organization divides work.

A person reads a situation and decides what it means. A policy narrows the available response. A software system records the decision. Another person reviews it if the consequence is large enough. A manager resolves an exception. An audit process returns later to ask whether the decision was reasonable. Each part contributes something different.

The person is useful where the world exceeds the schema. They can hear an unusual customer request, notice that two rules conflict, infer what an incomplete report means, or decide that a case should be escalated even when no field explicitly requires it. Human judgment is expensive and inconsistent, but it is adaptive.

The procedure is useful where variation is dangerous. It can require the same fields, apply the same calculation, preserve the same record, and refuse the same invalid transition every time. Deterministic systems are brittle outside their specification, but reliable inside it.

Organizations learned to combine the two. They did not ask people to improvise every payroll calculation, and they did not ask payroll software to decide whether an unusual payment was ethically justified. They placed flexible reasoning and rigid execution at different points in the same process.

This was not merely a division of labor. It was a division of epistemic responsibility. Some parts of the system were allowed to interpret. Others were expected to repeat.

The Employee Was Always a Bundle

It is easy to look at an AI agent and a human employee, observe that both can complete a task, and treat them as interchangeable actors. But the comparison isolates the ability to produce work from the institutional bundle that made the employee’s work usable.

An employee is not simply a source of intelligence. They enter through an identity. They occupy a role. The role grants authority and creates obligations. Their decisions can be reviewed. Their access can be revoked. Their reputation can change. They respond to incentives, professional norms, relationships, and possible sanctions. Someone can ask them why they acted, and the organization has established ways to accept, reject, or contest the answer.

None of this means accountability is a natural property stored inside the human body. Institutions create accountable people by placing identifiable humans inside systems of expectation and consequence. A signature matters because a role gives it meaning. A professional attestation matters because the profession can evaluate it. A manager’s approval matters because the manager has both decision rights and exposure to the outcome.

The person and the institution complete each other.

This is why replacing a human task with an agent action is not a simple substitution. The visible work may move cleanly. The surrounding bundle does not.

An agent can draft the recommendation, but who is obliged to defend it? It can execute the transaction, but whose authority makes the transaction legitimate? It can explain what it did, but what makes that explanation an accountable reason rather than another generated artifact? It can be retrained or shut down, but neither is equivalent to a person bearing responsibility for a decision.

The missing properties are not model capabilities waiting to be unlocked. They are institutional relationships waiting to be designed.

A New Material Enters the System

Traditional software is mostly governed by determining in advance what it may do. We define an interface, encode a rule, test expected behavior, and constrain its execution. The software may be complex, but the organization’s relationship to it is based on specification.

Human work is governed differently because it cannot be completely specified. Organizations use roles, training, supervision, professional standards, escalation, records, and retrospective review. The person receives discretion, but that discretion is surrounded by mechanisms that make it visible and answerable.

An AI agent sits awkwardly across these two traditions. It runs as software, but behaves less like a fixed pipeline. It can choose actions that were not enumerated by the developer, assemble tools in unexpected sequences, and interpret a broad goal in light of local context. Yet because it runs as software, we are tempted to govern it entirely through permissions and technical containment.

The result is a conceptual oscillation. When the agent appears intelligent, we treat it like an employee. When the question of responsibility arrives, we retreat and call it a tool.

Neither description is sufficient on its own.

The agent is a new material because it separates properties that used to arrive together. It gives us flexible reasoning without a human subject, scalable discretion without an existing professional role, and action without an obvious bearer of consequence. It is not institutionally unprecedented because it is intelligent. It is unprecedented because of the particular bundle it unbundles.

The Institution Is Between the Nodes

We often draw organizations as collections of boxes. A person, a team, an agent, a database, a service. The boxes appear to contain the important properties: intelligence, authority, data, capability.

But much of what makes an organization an institution exists on the lines between them.

A request becomes a delegated task. A draft becomes a reviewed document. An analysis becomes an approved decision. A technical result becomes an accepted risk. A private judgment becomes a public reason. These changes do not occur because the artifact moved from one inbox to another. They occur because the transition has an institutional meaning.

The handoff says who is passing what to whom, in what state, with which evidence, under whose authority, and with what remaining uncertainty. It may grant the next actor permission to proceed. It may require an attestation. It may preserve a record for later scrutiny. It may give the recipient the right to reject the work or escalate the decision.

The nodes perform the work. The edges make the work organizational.

This is visible in mature institutions. A junior lawyer can draft an argument, but a partner determines whether it leaves the firm. An engineer can propose a design, but an accountable authority accepts the risk. A payment initiator can create a transaction, but a second role releases it. A clinician can form a judgment, but the record, signature, review process, and professional duties make that judgment answerable.

These arrangements are sometimes dismissed as bureaucracy. Some of them deserve to be. But the general form exists for a reason. Institutions need mechanisms for turning local acts of interpretation into collective acts they can recognize, remember, and defend.

Accountability is not a property of the node. It is a property of the relationship between action, evidence, authority, and consequence.

How Judgment Becomes Trustworthy

Organizations have developed a family of mechanisms for governing flexible reasoners.

Delegation specifies which decisions belong to a role. Separation of duties prevents one actor from initiating, approving, and concealing the same consequential action. Escalation moves exceptions toward someone with different authority or context. Attestation makes an actor state that a condition was checked and accepted. Records make the action reconstructable after the moment has passed. Post hoc review tests individual decisions and reveals patterns that no single decision could show.

These mechanisms do not eliminate error. They do something subtler: they make judgment legible enough for an institution to work with.

A judgment becomes institutionally useful when others can determine where it came from, what evidence supported it, which authority accepted it, and what can happen if it proves wrong. The organization does not need to watch every cognitive step. It needs to preserve the structure necessary to evaluate the resulting act.

This is also why feedback loops matter. A review is not only a gate placed before action. It can be a sensor placed after it. A decision may proceed because it is reversible, then be sampled later for quality. A near miss may change a policy. Repeated failures may reveal that the problem is not an individual actor but a role definition, interface, incentive, or handoff that reliably produces the wrong outcome.

Institutions govern both prospectively and retrospectively. They constrain some actions before they happen and learn from others after they do.

The Human in the Loop Is Not an Institution

The usual response to agentic risk is to keep a human in the loop. The phrase sounds reassuring because it appears to restore responsibility to the system. But it does not tell us what the human is there to do.

If a person must approve every file read, shell command, and tool call, the human is not exercising judgment. They are servicing the agent’s execution loop. The first few requests may be inspected. Eventually the person learns that progress depends on pressing “Allow,” and the control becomes ritual.

A human click is not the same thing as human judgment.

The opposite arrangement is equally weak. An agent completes the entire process and a human signs the output without access to the evidence, uncertainty, or path by which it was produced. The signature attaches a person to the result, but it does not create meaningful review. It is responsibility without the conditions required to exercise it.

Human involvement becomes institutional only when the person occupies a defined role at a meaningful transition. They need a decision they are authorized to make, evidence appropriate to that decision, enough context to exercise discretion, and a clear account of what accepting the handoff will cause.

The question is not whether a human is somewhere in the loop. It is whether the loop contains a real institution.

The Platform Is Part of the Institution

Agent collaboration platforms are often described as orchestration systems. They route tasks, invoke models, provide tools, maintain memory, and coordinate several agents. That description makes the platform sound like neutral plumbing around the real intelligence.

It is not neutral.

The platform decides what counts as an actor, which identity appears in a log, what a role is allowed to do, when work is considered complete, which evidence survives a handoff, who may approve an effect, and when authority expires. Its schemas become claims about how the organization works. Its default workflow becomes a theory of responsibility.

If the platform models only agents and tasks, the missing institutional structure will reappear informally. Humans will pass context through messages. Broad credentials will substitute for scoped authority. Approval buttons will substitute for handoff policies. Responsibility will land on whoever happened to be watching when something went wrong.

An agentic collaboration platform therefore needs a richer set of objects. It needs agents, but also the roles they occupy. It needs harnesses, but also the habitats that bound their authority. It needs tasks, but also handoff policies that determine when work changes institutional status. It needs logs, but also evidence and attestations that make consequential transitions legible.

These are not safety features added around the edge of the platform. They are the material from which an agentic organization is made.

Designing the Relationships

The arrival of AI agents does not make the old organization obsolete. It reveals how much of the old organization was implicit.

We could treat agents as artificial employees and copy permissions designed for people. We could treat them as ordinary software and rely on containers, prompts, and API scopes. Both approaches preserve one side of the old division while losing the other.

The better approach begins with the relationships. Which role is delegating the work? What habitat defines the agent’s field of legitimate action? Which harness makes the work possible? What evidence must survive when the task moves? Which handoff changes a draft into a commitment? Who has authority at that junction? How can the decision be challenged, reversed, or learned from afterward?

The companion vocabulary is simple:

Role     = institutional position
Agent    = continuing actor
Harness  = machinery of agency
Habitat  = bounded authority
Handoff  = governed transition

The challenge is not merely to place agents inside the organization we already have. It is to redesign the relationships through which their actions become organizational acts.