Recommend reading this post - SaaS Repricing: What Everyone Gets Wrong About "Software Is Dead."

On Tuesday, February 3, 2026, approximately $300 billion in market value evaporated from software and data stocks. The proximate cause was Anthropic's announcement of specialized legal tools for its Claude Cowork assistant. But the actual cause—the thing investors were pricing—was something far more consequential than a new plugin.

The market was pricing in something the software industry had never considered: that every moat it built over twenty years was designed to stop humans, not agents.

The Catalyst vs. The Cause

The Wall Street Journal reported the losses in stark terms: Thomson Reuters down 15.8%, LegalZoom down 19.7%, PayPal down 20.3%. The selling spread from legal tech to payments, travel, and accounting. What should trouble anyone who thinks this was an overreaction is that the broader market barely moved. The S&P 500 fell 0.8%. The Dow dropped 0.3%. Five of eleven S&P sectors closed higher.

This was not a general panic. It was a targeted repricing of a specific thesis: that the application software layer, as we have known it for two decades, is being compressed out of existence.

The consensus view is that AI will compete with existing software by being better software. This is wrong. AI is not competing with the software layer. It is bypassing it entirely.

The Universal Interface

To see what changed, you need to know what Anthropic's Claude Cowork and OpenAI's Operator actually do. Previous AI tools were conversational interfaces—you asked a question, and they answered. The new tools are autonomous agents that operate at the desktop level, using a universal interface: the screen, the mouse, and the keyboard.

This is a major architectural shift. For 20 years, software companies built moats with proprietary interfaces and data silos. Users were trained on specific workflows. Data was locked into specific formats. Switching costs were high because switching required relearning and migration.

Agentic AI invalidates this entire defensive structure. An agent does not need to integrate with your API. It does not need to learn your proprietary workflow. It navigates your application the same way a human does—by looking at the screen and clicking buttons. The interface that was designed to create lock-in for humans creates zero lock-in for agents.

Anthropic's Model Context Protocol makes this worse. MCP lets agents connect directly to enterprise data systems, including Slack, Salesforce, Microsoft 365, and internal databases, creating a bridge between AI and the company's entire software stack. The agent becomes the integration layer. The individual applications become interchangeable commodities underneath.

This is not a feature improvement. It is a layer collapse.

The Microsoft Irony

The most telling detail from the week's coverage was buried in the Journal's report on Microsoft's Copilot struggles: "Claude Cowork has drawn praise for its ability to work across Microsoft 365 applications in ways that Copilot users find difficult."

Read that again. Anthropic's AI agent is better at navigating Microsoft's own software than Microsoft's own AI assistant.

This is not a failure of execution, though Microsoft has had many. It is a structural problem. Microsoft built Copilot to work within its applications, respecting the boundaries between Word, Excel, Outlook, and Teams. Anthropic built Claude to work across applications, treating the boundaries as obstacles to be navigated around.

Microsoft's approach preserves the value of individual applications. Anthropic's approach treats the applications as substrate: necessary but undifferentiated infrastructure that the agent manipulates to accomplish tasks.

The market understood this immediately. If the application layer is just substrate, application companies should be valued like substrate: at commodity multiples, not software multiples.

Where the Value Migrates

When a layer collapses, the value does not disappear. It migrates. The question for investors is: where?

Three destinations are emerging.

First, infrastructure. If agents need compute to run and data centers to host them, the hyperscalers win regardless of which applications survive. This explains why the selloff hit software stocks while leaving cloud infrastructure relatively unscathed. Microsoft's Azure business is fine. Microsoft's Copilot business is not.

Second, the model makers. OpenAI, Anthropic, and Google are becoming the arms dealers of the software wars. Every company that uses AI to attack a competitor is paying tokens to a foundation model provider. Every company that uses AI to defend its position is paying tokens to a foundation model provider. The model makers capture value on both sides of every transaction.

Third, trust and governance. This is the moat that remains. An AI agent can navigate a desktop and execute tasks, but it cannot provide the compliance certification a regulated enterprise requires. It cannot provide the audit trail that a public company's board demands. It cannot provide the contractual liability a risk-averse CIO needs before deploying mission-critical systems.

The companies that survive the layer collapse will be those that transform from software providers to trust providers, selling not the code but the governance, security, and accountability wrapper around it.

The Private Equity Canary

A less appreciated aspect of the February 3 selloff was the damage to private equity firms. Ares, KKR, and Blue Owl all dropped more than 9%. Apollo and Blackstone lost nearly 5%.

This matters because private equity has been the marginal buyer of software assets for a decade. The thesis was simple: software has recurring revenue, high retention, and predictable cash flows, the perfect collateral for leveraged buyouts. Blue Owl built a lending franchise around "recurring-revenue" software loans.

That thesis assumed switching costs were structural. If an AI agent can migrate data and rewrite integrations in minutes, switching costs are not structural. They are temporary. The "sticky" contracts that justified premium valuations become unstuck.

Jon Gray, Blackstone's president, acknowledged the problem at the Journal's investor conference: incumbent software companies that were "systems of record" now face "disruption risk" from AI. This is the sound of a marginal buyer reconsidering its thesis.

When the marginal buyer of an asset class starts hedging, prices have more room to fall.

The Eighteen-Month Test

The market's repricing is not complete, but it is directionally correct. The question is no longer whether AI disrupts software, but how far up the stack the disruption reaches.

The bear case is total layer collapse: agents handle everything from legal research to payment processing to travel booking, and the application vendors become pure utilities, valued at 1-2x revenue. This is the scenario implied by a 4.1x median revenue multiple.

The bull case is layer compression with value concentration: the middle of the market gets destroyed, but platforms with genuine data gravity and trust relationships, like ServiceNow, Workday, and the surviving CRM vendors, emerge stronger because they become the governance layer enterprises need to deploy agents safely. This is the scenario implied by ServiceNow's 98% retention and continued premium multiple.

The next eighteen months will determine which scenario prevails. The leading indicators to watch are not revenue growth or retention rates. They are:

  1. Expansion revenue trends. If existing customers are quietly negotiating smaller renewals while retention technically holds, the layer collapse is accelerating.
  2. Build vs. buy ratios. If CIOs start reporting that internal teams are building what they previously purchased, the addressable market is shrinking.
  3. Agent-based pricing adoption. If vendors successfully transition from per-seat to per-outcome pricing, they can capture value even as headcount declines. If they cannot, their revenue model breaks.
  4. MCP and protocol adoption. The faster Anthropic's Model Context Protocol becomes a standard, the faster the integration moat disappears for incumbent software vendors.

The Nature of the Shift

What happened on February 3 was not a panic. It was recognition.

For two decades, the software industry operated on an implicit assumption: that value accrued in the application layer. Users interacted with applications. Data lived in applications. Workflows were defined by applications. The company that owned the application owned the customer relationship.

Agentic AI inverts this. Users interact with agents. The agent interacts with applications. Applications become invisible infrastructure, necessary but no longer the locus of value or control.

This is not unprecedented. The browser did something similar to desktop applications in the 2000s. Mobile did something similar to the web in the 2010s. Each transition compressed the previous generation's moats and created new ones.

The difference this time is speed. Browser adoption took a decade. Mobile adoption took 5 years. Agent adoption, for those who want it, takes only a download and a subscription. The compression is happening faster than incumbents can adapt.

The $300 billion lost on a single Tuesday was not the market getting ahead of itself. It was the market catching up to a reality Silicon Valley engineers have discussed for months. The software layer is collapsing. The only questions that remain are how far it will fall and who will emerge from the rubble.

The sorting continues.