Headless software: your AI agent now picks the app for you

A line on Salesforce's product page has read the same way since April 15: "No browser required." The most visible vendor in B2B SaaS just stated that its platform no longer needs a human at the screen. The pitch is progress. The real question is: progress for whom?

SaaS goes headless

The announcement is called Salesforce Headless 360. The entire platform (data, workflows, business logic, compliance controls) is now exposed as APIs, MCP tools and CLI commands. More than 60 new MCP tools, 30 preconfigured "coding skills." The intended consumers are AI agents like Claude Code, Cursor or Codex, which can now drive the whole product without ever loading a browser.

Salesforce is not alone. Stripe runs an official MCP server at mcp.stripe.com. Shopify has shipped four (Storefront, Customer Account, Checkout, Dev). Twilio plugs in through Zapier MCP, itself wired to 9,000 integrations. The Model Context Protocol, launched by Anthropic in late 2024, has become the de facto standard in eighteen months for any agent that needs to talk to a piece of software.

Aaron Levie, CEO of Box, has been blunt about it on X: "Agents are going to use software 100x more than humans. Enterprise platforms will go headless and work with any agent. If you don't do that, you're dead." Translation: the graphical interface stops being the product. It becomes a legacy option for the humans still hanging around the screen.

A pivot pushed by interested parties

Before swallowing the narrative whole, a useful caveat: Salesforce, Box and Stripe each have a direct stake in pushing the "everyone goes headless" scenario. More API calls means more revenue for Stripe. Levie has publicly floated a model where every agent pays a Box seat, on top of human seats. Salesforce, on its own TDX 2026 slides, talks up a platform that runs "wherever you work." The more usage shifts to agents, the more value the vendor captures.

Gartner reckons vendors that simply bolt AI onto legacy stacks, without redesigning for agentic execution, face up to 80% margin compression by 2030. The figure is being passed around as a sales argument: "redesign now or die." Maybe. But the prophecy is also conveniently aligned with the interests of the people selling the transition.

What changes for you, the user

This is where the story leaves the B2B sandbox and touches everyone. Today, picking an app feels like picking a car: looks, ergonomics, habit. Tomorrow, the pick is an agent. The apps it touches become interchangeable back-ends you will never see.

A concrete example: you ask your assistant to book a Paris-to-Lyon trip. Today it opens an app you know, you see the interface, you check the price. Tomorrow it queries three MCP servers (SNCF Connect, Trainline, any aggregator), picks the cheapest based on criteria it inferred from your past behavior, and books. The screen only shows the confirmation. The brand of the app used drops out of frame.

Two practical consequences for the broader public:

First, the quality of an app stops being something you can see. A polished UI, slick marketing, a clean onboarding flow, none of that carries weight if an agent is consuming the app. What matters becomes invisible: API documentation quality, MCP server latency, accuracy of the data returned. The vendor can quietly degrade the human-facing interface and nobody notices, because nobody looks at it anymore.

Second, the stickiness moves. Apps that used to lock users in through habit and UI become substitutable. The agent, on the other hand, accumulates behavioral capital: it knows your preferences, your exceptions, your decision patterns.

Analysts call this behavioral lock-in. And that lock, unlike data, does not transfer. European portability rules (DMA, GDPR) only cover the explicit. Learned behavior stays trapped inside the agent that observed it.

Consumer adoption stays slow

Whether this scenario arrives next year or in five years remains an open question. On the consumer side, autonomous agent penetration is moving slower than the narrative suggests. The Prophet 2026 report counts 73% of users on GenAI (up from 45% two years earlier), but only 13% have completed a purchase based on an AI agent's recommendation. Most usage stays in conversation and search, not autonomous delegation.

Apple, Google and Microsoft control the consumer gateways (operating systems, browsers, productivity suites) and none have announced any retreat from human-facing interfaces. Apple is even floating an option to let users plug in their own third-party AI model through "Extensions" planned for iOS 27. The "everything routes through the agent" pivot is real on the B2B side, but on the consumer side it is still a bet.

The shift is far enough along, though, that the right question starts changing. Not "which app do I use?" but "which agent gets the keys?" Because the agent is the one picking everything else.

As US analyst MindStudio put it in a May 2026 study: "The real asset of the agentic era is not data, it is behavioral context. And behavioral context is not portable."