ON SYSTEMS OF ATTENTION

Systems of Attention: The Missing Layer Underneath the AI Agent Crisis

Microsoft restructured Build 2026 around it. Gartner named it agent sprawl. The security industry is drowning in it. The cognitive research has the numbers. Five witnesses to the same missing layer of enterprise software, and the vocabulary is finally catching up.

Five recent tier-one stories and news items. Different vocabularies, same diagnosis. The enterprise AI agent crisis is not five different problems. It is one missing layer of enterprise software, and the industry is learning its name.

The enterprise AI agent crisis is real, and it has been documented in five different vocabularies across the past two weeks. Microsoft restructured its flagship developer conference around it. Gartner gave it a name. The security industry is drowning in its absence.

The cognitive research has the numbers to prove what every operator already feels. Stand at the right distance and they are all describing the same missing layer.

Something happens when you read two weeks of enterprise AI coverage in one sitting instead of one story at a time. The threads stop looking like separate stories. They start looking like five witnesses describing the same event from five different rooms in the same building. Each one is telling the truth about what they saw. None are quite far enough back to see the whole room.

So rather than respond to any one piece, I want to walk through what showed up across the coverage and then say what the industry is actually circling. The diagnosis is converging fast. The vocabulary is not.

What Each Witness Saw

Start in San Francisco. On June 2, Satya Nadella stood on the Fort Mason stage in front of five thousand developers and made a declaration that will be quoted for years. "Agents are not just a feature," he said. "They are the new operating system for work."

Microsoft Build 2026 had been positioned as a developer conference. It turned into something larger. Microsoft Agent Framework 1.0 reached general availability. Office 365 Copilot Agent Mode shipped. Microsoft Execution Containers were introduced as a new policy layer for keeping autonomous agents bounded. The throughline across every track of the conference was the same: agent orchestration, agent governance, the question of how enterprises will keep up with what they have already deployed.

You do not build the world's largest developer event around a solved problem. You build one around the problem the market has surfaced faster than the architecture has answered.

Five weeks earlier, at the Gartner Digital Workplace Summit in London, Senior Director Analyst Max Goss presented a number that has been traveling through CIO conversations ever since. By 2028, an average global Fortune 500 enterprise will run over 150,000 AI agents. Up from fewer than 15 in 2025.

Gartner now has its own vocabulary for what comes next, and the vocabulary is sharp. They call it agent ‘sprawl.’ Goss put the framing plainly. CIOs are contending with an ungoverned proliferation that exposes their organizations to misinformation, oversharing, and data loss. Blocking the tools, he said, only pushes employees toward shadow AI, which carries even greater risk.

Step out of the analyst conversation and into the Security Operations Center (SOC). Current research from the State of AI in Security Operations survey, polling nearly three hundred CISOs and practitioners, puts the median security team at about 960 alerts per day.

• Roughly 40 percent are never investigated.

• Analyst burnout sits at 71 percent.

• Average tenure has compressed to three to five years before the role chews someone up.

Older Forrester research from 2020 placed the high end above 11,000 alerts daily, and that number is still cited because it remains directionally accurate for the largest enterprises. The industry calls it alert fatigue. Every practitioner who has worked a SOC desk names the same root cause underneath. It is not volume. It is the absence of context. An alert without context is noise. A wall of alerts without context is a wall of static no human ear was built to parse.

Step out of the security floor and into the cognitive evidence. The research published earlier this year by Boston Consulting Group and UC Riverside in Harvard Business Review put a name on what engineers under heavy AI agent use have been quietly describing. They called it ‘AI brain fry.’ A study of 1,488 full-time U.S. workers found 14 percent reporting acute cognitive overload, with 33 percent more decision fatigue and 39 percent more major errors than peers with lower oversight load. The research is several months old now, but it remains the only quantified evidence of the human cost, and the numbers have aged into the floor of the conversation rather than out of it.

Picture what it looks like in practice. An engineer with five agents running before lunch. Each one finishing a task and pinging for review. Accept, edit, reject, escalate. Multiply that by a hundred decisions a day on top of the actual work.

The researchers were careful about the punchline. The damage was not caused by using AI. It was caused by overseeing it.

Then there is the question of what the enterprise actually has running inside its walls right now. IBM's Think conference data, presented earlier this spring and still moving through CIO conversations, projects most large enterprises will operate over 1,600 AI agents by year-end. Seven in ten executives told IBM their current AI governance is not fit for purpose. Only 18 percent can produce an inventory of the agents already running.

Eight days ago, just before Build opened, Microsoft's own Copilot Studio update moved three critical capabilities out of preview into general availability. Computer-using agents that interact directly with the UI layer of enterprise software. Agent-to-agent communication for delegating subtasks. Real-time voice. The Microsoft Agent 365 enterprise control plane reached general availability on May 1. None of these capabilities existed at this maturity level a quarter ago. All of them shipped in a single window.

What These Witnesses Are Naming

Read those threads back-to-back. The largest software company on Earth restructuring its developer conference around agent orchestration. Gartner naming agent sprawl as its own analytical category. SOC teams drowning in 960 daily alerts that 40 percent of the time never get a human's eyes. Engineers under heavy AI oversight showing measurable cognitive overload. Most large enterprises about to operate more than 1,600 AI agents they cannot inventory.

Five witnesses. Five rooms. Same event. None of them used the same word. All of them described the same thing.

The Unnamed Primitive

What the past stretch of coverage is collectively pointing at is a layer the enterprise stack has never actually built. It’s not a feature or a tool, but a structural primitive that has been sorely needed over the last several decades, and that AI has brought to the surface.

Enterprise software has evolved in three layers.

Systems of Record came first, capturing what happened. Think SAP, Oracle, and Salesforce.

Workday, ServiceNow, and the Saas era ushered in Systems of Engagement: mediating how people interacted with the record.

Then Systems of Insight, surfacing patterns across the data. The modern data stack from leaders such as Snowflake and Databricks.

Each layer made a different assumption about the layer above it. All three made the same assumption about the human at the end of the pipe. That person who would do the noticing. The triage. The judgment about which signal mattered most in the key moment to act.

That assumption was already under strain, in real-world business challenges and poor decisions, before AI got cheap. It is broken now, and the past several weeks of stories has been revealing. When a single developer has to monitor a dozen agents, brain fry is the inevitable result. When an enterprise runs 1,600 agents and cannot inventory them, governance becomes theater. When 960 alerts hit a SOC every day, the human can no longer be the arbiter of which ones matter. When Microsoft introduces Execution Containers as a new policy layer for keeping agents bounded, the company is acknowledging architecturally what the rest of the industry is naming symptomatically.

The architecture asked humans to do attention work at a volume and velocity that humans cannot perform. Not because humans got worse. Because the input rate ran past the cognitive ceiling, and nothing in the stack ever picked up the difference.

Enter Systems of Attention: The Fourth Layer of Enterprise Software

Systems of Attention is the fourth layer of the enterprise stack. It sits above Record, Engagement, and Insight. It is the architecture for converting signals into routed decisions, for treating human attention as a finite institutional resource, and for closing the loop so the next instance of a pattern does not start from scratch.

Underneath the layer are four primitives that make it real:

1. Signal Ingestion: where raw data becomes signal nodes with scoring and decay rates
2. Equilibrium engine: holds a working read on normal so the abnormal surfaces.
3. Attention Routing: ranks signals by criticality and within quantity limits
4. Decision Memory: every decision writes writes back, where the system learns and compounds.

Decision intelligence platforms address part of this. They model decisions, execute them, and govern their outcomes once a decision has been chosen to make. But the layer above decision intelligence, the one that decides which signals to inform decisions earn the institution's attention in the first place, has not existing as a category.

That is the layer the industry is now circling. Enter Systems of Attention: the vocabulary has not caught up to the diagnosis.

Naming the layer matters because the diagnosis determines the response. Read the past stretch of coverage as five disconnected problems and you will buy five disconnected solutions.

• A wellness program for brain fry.

• A governance tool for agent sprawl.

• An orchestration platform for coordination.

• A SIEM tuning project for alert volume.

• A change management initiative for AI rollout.

Each of those addresses one symptom. None of them address the structural fact that the enterprise has no architectural layer responsible for deciding what deserves attention right now, who should see it, how confident the read should be, and what to remember when the moment closes.

What CIOs and CSOs Reading This Coverage Should Ask Next

If you are a senior technology leader watching this pattern resolve, three questions are worth taking into next week's leadership meeting. Each one is simple to ask and harder to answer than most enterprises expect.

What is the attention budget per executive, per analyst, per operator? Most enterprises cannot produce a number. The right number is not infinite. The decision-fatigue and cognitive-load research converges on the same finding. Human attention is a finite, depletable resource. Treating it as if it scales linearly with software output is exactly what got the industry to brain fry.

Where in the stack does institutional memory actually live? When an incident closes, when a decision is made, when a signal gets routed correctly, does that learning write back to anything? Or does the next instance of the same pattern start from scratch? Most enterprises will find the latter is true, and the cost compounds quietly until the next crisis exposes it.

When the agent count goes from 30 to 300 to 1,600, what is the architectural primitive that decides which agent's output deserves a human's attention right now? If the answer is the same dashboards and inboxes the enterprise has always used, the brain fry research has already told you how that ends. The architecture cannot absorb the load.

Underneath all three questions is the same uncomfortable truth. The reason the recent coverage feels so dissonant is that the industry has been describing a structural problem with the vocabulary of individual symptoms. Cognitive overload is real. Alert fatigue is real. Agent sprawl is real. Governance gaps are real. None of those are the root condition. They are the surface expressions of a single missing layer in the architecture, and they will keep multiplying as long as that layer remains unbuilt.

The stories from this past stretch are not separate problems. They are one problem photographed from different angles by witnesses who never compared notes. The resolution only becomes visible when you stop reading them one at a time and start reading them as a single pattern.

The AI agent crisis is not five different problems. It is one problem photographed from different angles by witnesses who never compared notes. Systems of Attention is the layer underneath. The fourth layer of enterprise software is the architectural claim. Building it is the work of the next decade. The vocabulary will follow.

Sources Referenced

1. Microsoft Build 2026 (June 2-3, 2026): Be Yourself at Work. Official Microsoft Blog.

2. Gartner Identifies Six Steps to Manage AI Agent Sprawl (April 28, 2026). Max Goss, Senior Director Analyst, Gartner. Gartner Digital Workplace Summit, London.

3. IBM Think 2026: Shaping the next era of agentic AI (May 4-7, 2026). Andy Baldwin, SVP Offerings and Growth, IBM. IBM Think Conference, Boston.

4. When Using AI Leads to "Brain Fry" (March 5, 2026). Julie Bedard, Matthew Kropp, Megan Hsu, Olivia Karaman, Jason Hawes, and Gabriella Kellerman. Harvard Business Review. Boston Consulting Group and UC Riverside research.

5. State of AI in Security Operations Survey (2025-2026). Prophet Security. Nearly 300 CISOs, SOC leaders, and practitioners.

6. Forrester Consulting State of Security Operations (2020). Commissioned by Palo Alto Networks. Source for the 11,000-alerts-per-day high-end estimate.