Architectures of Coordination
Agents, information systems, and the firm, with Jack Dorsey's essay as a starting point.
Jack Dorsey's recent essay, From Hierarchy to Intelligence , is interesting not mainly as a claim about AI, but as a claim about the architecture of the firm. Its central argument is that hierarchy has historically functioned as an information-routing mechanism, and that new computational systems may allow organizations to coordinate in ways that rely less on managerial layers and more on machine-maintained models of reality.
This is a serious proposition, and it deserves a more serious treatment than the usual language of AI transformation. The important question is not whether AI will replace managers, nor whether organizations will become intelligent in some vague promotional sense. The question is architectural: what kinds of information systems would be required for agents to participate meaningfully in coordination, and what does this imply about the design of the modern firm?
Dorsey is right to begin with hierarchy. The classical hierarchy solved a real problem. Organizations are not merely groups of people with shared goals, they are systems that must continuously gather signals, interpret changing conditions, allocate attention, make decisions, and coordinate action under uncertainty. Historically, hierarchy has been one of the most effective mechanisms for doing this at scale. A chain of command is also a chain of information aggregation. Reports move upward, decisions move downward, and each managerial layer serves as a compression and translation point.
This arrangement was never philosophically elegant, but it was practically robust. It matched human cognitive limits, communication constraints, and the need for accountability. Modern management structures, for all their frustrations, persist because they are not arbitrary. They are adaptations to the difficulty of coordination in complex environments.
What makes the present moment different is not simply the existence of better automation. It is the emergence of systems capable, at least in principle, of participating in the maintenance of organizational state. If a sufficiently rich stream of organizational activity becomes machine-readable, then parts of the work previously carried by managers, project leads, coordinators, and operational staff might be redistributed into information systems. But this depends on architecture, not on rhetoric.
Agents cannot coordinate around what they cannot perceive. They cannot reason over what is not represented. And they cannot act reliably in environments where organizational state exists only as fragmented documents, tacit knowledge, and transient chat.
For this reason, the key issue is not whether models are smart. It is whether the firm can produce a sufficiently legible and operational representation of itself.
From Documents to State
This requires a shift from document-centric systems to state-centric systems.
Most organizational tooling today is artifact-heavy but state-poor. Companies produce tickets, documents, messages, dashboards, recordings, code commits, spreadsheets, and meeting notes in enormous volume. But these artifacts do not automatically form a usable model of the organization. They are traces of work, not a coherent account of what is currently true. An agent does not benefit much from being given access to a large pile of documents if the relevant structure, status, dependencies, and authority relations remain implicit.
A more agent-compatible architecture would require at least four things.
Goals, plans, tasks, decisions, owners, constraints, risks, dependencies, and statuses must exist in structured form, not only as prose scattered across systems. This does not mean reducing the firm to a database schema. It means acknowledging that coordination requires explicit state that can be updated, inspected, and contested over time.
Organizations are not static repositories, they are streams of changes. A priority shifts, a customer escalates, a deployment fails, a requirement is clarified, a budget changes, a dependency slips. If agents are to participate meaningfully, they need architectures built around events and updates, not merely periodic snapshots or manually prepared summaries.
Once agents begin contributing to analysis or action, it becomes important to know what context they used, what assumptions they relied on, what actions they proposed or took, and what state they modified. Without traceability, automation may accelerate activity while degrading accountability.
Not every signal should be treated equally. A Slack message, a customer payment, a board decision, a failing test, and a speculative comment do not have the same epistemic status. Effective organizational systems need ways to represent trust, provenance, recency, ownership, and domain relevance.
The World Model, Treated Carefully
Seen from this perspective, the notion of a world model should be treated carefully. The term is useful, but it can also be misleading. A company does not possess a perfect model of itself, nor can it. Organizational reality is incomplete, disputed, and temporally unstable. Different actors hold different interpretations of the same situation. Incentives distort reporting. Metrics conceal as much as they reveal. Strategic priorities are often underdetermined. The problem is not to construct a final and objective representation, but to maintain a revisable and operationally useful one.
This is why planning remains important, though perhaps not in the usual sense. Plans and tasks are not the whole firm, but they provide one of the most practical interfaces between perception and execution. They create operational surfaces on which intentions, dependencies, responsibilities, and progress can be represented in a form accessible to both humans and agents. In this sense, planning systems may become part of a broader coordination architecture, one that links raw signals to interpretable state and interpretable state to action.
Human-Agent Coordination, Not Managerial Erasure
That architecture should not be imagined as a system for replacing human judgment wholesale. The more plausible and more desirable model is one of distributed coordination between humans and computational agents. Agents may assist in maintaining state, surfacing inconsistencies, proposing next steps, monitoring dependencies, or routing information. But organizational interpretation remains a contested and normative activity. What matters is not only what is true, but what matters, what should be prioritized, what tradeoffs are acceptable, and whose judgment should prevail under uncertainty. These are not problems that disappear when information systems improve.
Dorsey's essay is valuable because it directs attention to the right level of analysis. The future of organizations will not be determined merely by better copilots or by isolated task automation. It will depend on whether firms can redesign their coordination architecture around machine-readable state, event-driven information flows, and systems that support both computational inference and human oversight.
This is a more demanding challenge than the current AI discourse often suggests. It is not mainly a question of interface design or prompt quality. It is a question of institutional information design. What must be represented? What should be updated automatically? What needs review? What counts as a trusted signal? How should conflicts between representations be resolved? What forms of memory should persist, and at what level of abstraction? These are architectural questions, but they are also questions about management, epistemology, and the theory of the firm.
If there is a real shift underway, it is here. The firm may increasingly be understood not just as a hierarchy of roles or a nexus of contracts, but as an evolving information system that must maintain a usable model of itself while acting in the world. In such a setting, the role of AI is not simply to generate outputs faster. It is to participate in the representation, revision, and coordination of organizational reality.
That possibility is substantial. But it will only matter if organizations build the underlying architectures that make it possible.
Working on coordination systems for agents?
We design planning, memory, and execution layers for real human-agent workflows.