For two decades the home screen was the center of gravity. Users tapped your icon, your native app opened, and you owned the session. That contract is dissolving. The next entry point is not an icon — it is an agent. And most traffic apps, having pushed their core flows into WebView remote documents, are structurally unprepared to be driven by one.

The launcher icon is losing its monopoly

Google did not ship a search engine update in 2025; it shipped a different product. AI Mode, rolled out globally in May 2025, answers in a conversational interface where roughly 93% of searches end without a click to any external site (77% on mobile). Where an AI Overview appears, organic click-through drops about 61%, according to Seer Interactive's 25-million-impression study. Global publisher traffic from Google fell roughly a third year over year to November 2025 (Press Gazette / Chartbeat). Gartner projects traditional search volume to decline 25% by 2026.

Apple is closing the same loop from the other side. App Intents lets any app expose features to Siri's reasoning layer — «find the invoice I was working on and send it to yesterday's contact.» The agent, not your icon, becomes the dispatcher. OpenAI, Google, and Microsoft have all standardized on the Model Context Protocol (MCP) to let agents reach tools and data; Anthropic's spec, open-sourced in November 2024, hit 97 million monthly SDK downloads and 10,000+ active servers by December 2025 and was donated to the Linux Foundation.

The pattern is uniform: the interface that decides what a user does next is migrating from the launcher to an agent that can perceive, decide, and act.

What an agent needs to drive your app

An agent that books the ride, completes the purchase, or answers the query needs one thing a WebView cannot reliably give it: a stable, structured, first-party capability interface. «Show size 43, add to cart, apply the loyalty discount, check out» is a deterministic API call. It is not a freshly fetched HTML blob whose buttons moved since yesterday.

Three properties matter:

  • Discoverability — the agent must learn what your app can do from a contract, not by parsing a live DOM.
  • Determinism — the same call yields the same result, so the agent can plan and verify.
  • Context continuity — the session survives, so a half-finished task is not lost when the network blinks.

Why WebView fails the agent test

A WebView screen is a document fetched from your server on every session. To an agent it is nearly opaque:

Property an agent needsWebView remote documentFirst-party runtime component
Discover what the app can doParse a changing live DOM — no stable contractVersioned capability API, declared up front
Drive a flow deterministicallyScrape and click brittle selectorsCall a named, signed action
Keep context across the sessionRe-fetch drops state; storage is isolatedSession owned by the runtime, persists
Survive weak networksBlank spinner until the doc reloadsRuns locally once packaged
Stay inside your security boundaryJS bridge reaches native on compromiseSandboxed, capability allow-list
Roll back a bad changeRe-push to 100% or nothingVersioned package, staged rollback

The same fragilities we have documented across this series — the discoverability black hole, the broken handoff, the offline collapse, the exposed attack surface — all converge here. An agent is simply the strictest possible consumer of your app, and WebView fails every strict test.

Agentic commerce is already redirecting the funnel

This is not hypothetical demand. KPMG's Consumer Snapshot 03/2026 found 46% of consumers used AI frequently in purchasing decisions over the prior twelve months, and 73% can imagine using an agent to compare products, prices, or offers. Morgan Stanley forecasts nearly 50% of online shoppers will use AI agents by 2030, accounting for about 25% of their spending and roughly $115 billion of US e-commerce. Adobe Analytics shows AI-driven traffic to US retail surged 805% year over year on Black Friday 2025, and in Q1 2026 AI-referred shoppers converted 42% better than non-AI traffic — a full reversal from a year earlier.

The catch is the one this series has hammered from the start: an agent does not read your marketing copy. It reads structured data — price, availability, delivery, return policy, product attributes. If that data lives only inside a remote WebView document with no machine-readable contract, the agent cannot see you, and you simply do not exist in that transaction.

The runtime answer: make your app callable

The fix mirrors what the industry did with MCP. Package your mini-program as a first-party, signed, sandboxed component that exposes a versioned capability API — an «MCP server for your app.» Your own agent, the platform's agent, or a third party's can call it deterministically. The runtime renders natively, keeps the session, and you own the contract and the rollback. You keep the entry point instead of ceding it to whoever controls the agent.

On-device AI makes this sharper. Apple Intelligence and Google's Gemini Nano run locally on Neural Processing Units; the dominant 2026 pattern is on-device first, cloud fallback. An agent that lives on the device needs a local, first-party capability surface to act inside your app — exactly what a runtime component provides and a remote document does not.

The home screen did not disappear. It was promoted. The apps that survive the agent era will be the ones an agent can actually call.

Sources