The Queue Was the Operating Model

Its visible objects are alerts, tickets, bridges, runbooks, change windows, dashboards and SLAs. Its invisible object is the containment of uncertainty.

When something breaks, the enterprise does not simply repair a system. It summons an organization.

This is why incidents feel so expensive even when the technical fix is small. A failed service may require a restart, but the enterprise requires triage, ownership, evidence, permission, communication, escalation, validation and memory. The technical event may be minor. The coordination event around it is often large. Modern IT operations is the practice of surrounding machines with enough human ceremony that failure becomes administratively manageable.

I.

Ronald Coase understood something management theory often forgets; organizations exist because coordination has a price. Markets are not free to use. Every transaction carries the cost of search, comparison, negotiation, monitoring, enforcement and correction. When that cost becomes too high, work is pulled inside the firm. The firm is not the opposite of the market. It is the market’s friction made institutional.

The firm grows until its own internal bureaucracy becomes as expensive as the external friction it was built to escape. That is the Coase equilibrium. It is not a preference for scale or a romance of the corporation. It is a boundary condition. Work settles where coordination is cheapest, safest and most controllable.

Enterprise IT ops follows the same law. NOCs, SOCs, CABs, L1/L2/L3 towers, escalation matrices, vendor desks, ITSM queues and command bridges are not accidental bureaucracies. They are the institutional form of operational transaction cost. The enterprise built them because, at scale, even knowing who should act becomes a problem.

II.

Autonomous agents disturb this equilibrium. The interesting part is not whether they replace engineers. That question is too small. The interesting part is that they reduce the cost of coordination itself.

An agent can read the alert, inspect the telemetry, correlate the change, search incident history, query the CMDB, update the ticket, execute the runbook, verify recovery and preserve the audit trail. None of these actions is miraculous in isolation. The significance is that the handoffs collapse.

The old path was human-mediated -> alert, ticket, queue, escalation, bridge, action. The agentic path is evidence-mediated -> signal, context, policy, action, verification, audit. This is the architectural shift. The agent is not a chatbot sitting beside operations. It becomes a coordination layer between signals, systems, policies and consequences.

III.

AI bends the Coase curve inside IT operations. In the old model, operational scale made human coordination more expensive. More systems created more owners. More owners created more queues. More queues created more latency. More latency created more process. More process created less actual understanding.

The enterprise often called this maturity. In many cases, it was only accumulated friction with better vocabulary.

Agents flatten part of that curve by moving repeatable operational work out of human queues. The work does not disappear; it stops needing human passage. A failed pod, a disk threshold, a duplicate alert storm, a known rollback, a recurring batch failure, a deployment-correlated outage; these do not become meaningful work because a human is forced to touch them.

Much of IT ops is not judgment. It is coordination residue. Residue should be mechanized.

IV.

The first domain of autonomy is not mysterious. It is the domain where the signal is known, the diagnostic path is known, the action is bounded, the blast radius is low, rollback exists and the outcome can be verified.

This is not science fiction. It is the removal of work that should never have required human coordination in the first place.

The enterprise should not begin with “autonomous operations” as a declaration. It should begin with operational loops that can prove they no longer deserve a human queue. Autonomy is not a capability one buys. It is a permission earned through evidence.

V.

The easiest interpretation will be headcount reduction. Every serious technology is first translated into the language of cost. The argument will be simple; if agents reduce operational effort, fewer people are needed. That may sometimes be true, but it is the least interesting version of the story. It is also how fragile enterprises make themselves cheaper before making themselves weaker.

The deeper implication is that humans should move closer to reliability, architecture, governance and risk. The human stops being a biological router for tickets and becomes the designer of permissible autonomy.

The question changes from “who can fix this?” to “what is the system allowed to fix by itself?” That question is not purely operational. It concerns authority, boundary, accountability and trust.

VI.

This is where the naive AI story breaks down. Agents do not eliminate transaction costs. They relocate them.

The old costs were queues, handoffs, escalations, bridge calls, manual triage, runbook hunting, vendor chasing and tribal memory. The new costs are identity, permission, policy, evidence, telemetry trust, rollback, auditability, blast-radius control, liability and compliance.

The amateur question is whether the agent can perform the task. The enterprise question is whether the action can be trusted, bounded, explained, reversed and owned. Intelligence is not the scarce commodity in production operations. Consequence is.

An agent without governance is not autonomy. It is automation that has learned to speak with confidence.

VII.

The right architecture is not “agents everywhere.” That is how vendors describe the world when they want to avoid boundaries. The useful architecture is bounded agents inside a governed operational control plane. The model is not the system. The system is the discipline around the model: identity, policy, observability, action, verification, rollback and audit.

This is not an academic distinction. Enterprises rarely fail because nobody could imagine the right action. They fail because permissions were too broad, ownership was unclear, evidence was weak, telemetry was polluted, rollback was absent, exceptions were unmanaged or accountability disappeared into the operating model.

The future of IT ops will not belong to the cleverest agent. It will belong to the safest useful autonomy.

VIII.

Autonomy has to be staged because trust has to be staged. First the agent observes. Then it explains. Then it recommends. Then it executes low-risk reversible actions. Then it performs governed changes through policy, approval, canary and rollback. Only after this progression does the language of autonomous resilience become respectable.

Many enterprises speaking loudly about full autonomy have not yet mastered explanation. That is not a model gap. It is an operating model gap. They want the machine to act before the organization has learned what action should mean.

IX.

For managed service providers, this is not a feature extension. It is a challenge to the economic logic of the industry.

The old MSP model monetized coordination friction. More alerts became more tickets. More tickets justified more people. More people required more shifts. More shifts produced more SLA reporting. More activity became more contract value.

Agentic operations attacks the center of that arrangement. Triage, correlation, ticket updates, routine remediation, incident timelines, reporting and runbook execution are exactly the activities most exposed to agents. The labor pyramid starts to look less like an operating model and more like a queue with contractual protection.

The new MSP cannot simply sell people who manage tickets. It has to sell governed autonomy. The unit of value moves from effort to evidence, from staffing to resilience, from activity to avoided coordination. That transition will be hard for an industry trained to invoice friction.

X.

The Coase equilibrium in IT ops is moving. Human ownership remains where trust, risk, liability, architecture and judgment matter. Agents take what is repetitive, observable, reversible, measurable and policy-bound.

The old model was humans coordinating machines. The new model is humans governing agents coordinating machines. This changes org design, tooling, commercials, managed services and the meaning of operational excellence itself. The winners will not be the enterprises with the most agent demos. They will be the ones with the clearest boundaries; where autonomy belongs, where it is forbidden, where humans must judge, where machines may act and where evidence must decide.

The future of IT ops is not autonomous everything. That is a child’s idea of intelligence. The future is governed autonomy. And the enterprise that understands this will not merely reduce cost. It will carry less organizational weight, move with less ritual and recover with less theatre. It will not celebrate how many incidents humans resolved. It will ask the more severe question.

Why did this incident require human coordination at all?

Originally published on LinkedIn