Who gets paid
Question two: who gets paid?
The interface decision is also a business decision, and each frame makes money differently. One warning from startup investors is worth taking seriously: the moats left are distribution, proprietary data, and services, not the interface you pick (post). So the question isn't only where your software meets people. It's who owns the customer when it does.
Skills rent land. A skill lives inside someone else's agent, and that platform is your landlord. It controls discovery, ranking, and often the customer relationship. The upside is distribution you could never buy: ChatGPT recommending plugins to 1.2 billion weekly users. The downside has a long history. Apple has taken a 30% commission on paid apps and in-app purchases for most developers, and a federal court found it in contempt in 2025 for blocking developers from linking to cheaper payment options and charging a 27% fee on purchases made outside the app (MacRumors). Zynga built an empire on Facebook, then disclosed in its 2012 annual report that substantially all of its revenue came through Facebook, which kept 30% of Facebook Credits purchases. Platforms give you reach, then set the terms.
The defensible skill has a paid core. A skill that's only clever instructions is easy to copy, by competitors or by the platform. The instructions can be free. What gets paid for is what's behind them: proprietary data, a live service, a model or algorithm you run, a human expert in the loop. The new commerce rails are built to let that core keep its customer: under the Agentic Commerce Protocol, the merchant accepts the order, takes the payment directly, and handles fulfillment. My trainer's prompts are portable, but a scoring engine that tracks how your reasoning improves over months could sit behind a subscription the skill calls. Think of the skill as the storefront and the service as the business.
Traditional apps own the relationship. An app has become less expensive to build but is still harder to get people to open. In exchange, the customer, the habit, and the data are yours (minus whatever the app store takes). That's the real reason to build one, beyond the presence test: it's the frame where you come closest to owning the relationship with your customer.
Ambient apps win on trust. Software that sees your screen and hears your calls is only valuable if people trust it with everything. Clicky's pitch leads with that: it only looks at your screen when you press the hotkey, and it says it never stores screenshots. Investors have already flagged privacy concerns around smart glasses (Reuters). The ambient businesses that last will charge for trust as much as capability, and they're the most exposed to the operating systems that want to do the same thing natively.
Agents become customers. This is the shift most founders miss. Agents won't just carry your software to people; they'll buy it themselves. Some investors already size building for agents as buyers as a $100B+ opportunity (post), and the payment rails are already shipping:
Google announced the Agent Payments Protocol (AP2) in September 2025, an open protocol for agent-led payments. Coinbase's x402 revives the web's long-unused "402 Payment Required" status code so an agent can pay per API request, with no account, API key, or invoice. Coinbase says those stablecoin payments can settle in under 500 milliseconds for less than a tenth of a cent (Coinbase). Per a MeshPay roundup citing Chainalysis and the Linux Foundation, x402 passed 100 million transactions on Base and now sits under a Linux Foundation group whose members include AWS, Google, Stripe, Visa, and Mastercard. Visa's Intelligent Commerce, which gives agents tokenized cards with user-set limits, is live in the US, according to the same roundup.
That makes the paid core of my trainer concrete: an agent could pay a few cents per scoring call to a service that sits inside a free skill. Building for agent customers means clean APIs, machine-readable pricing, per-call billing, and reliability over polish, because an agent doesn't care about your onboarding screens.
The forecast, and the signals to watch
I see the unbundling happening in three stages. They overlap, and each comes with signals that would tell me I'm right or wrong.
1. Software moves into the chat box. Capabilities become skills, and apps that were secretly ways of asking collapse into conversation. Evidence: apps and checkout inside ChatGPT, a skills standard adopted across coding agents, and 10,000+ MCP servers. Signal to watch: plugin and skill directories becoming a real acquisition channel, with agents recommending tools unprompted and those referrals showing up in founders' revenue.
2. The chat box moves into the world. Through voice, cameras, and glasses, the assistant leaves the window. Evidence: production speech-to-speech models, assistants that can see through your camera, and 7 million AI glasses sold in 2025. Signal to watch: voice becoming the main way people use a vertical product, not a demo feature, and AI-glasses growth holding up after this year's slowdown.
3. The agent moves into the operating system. Agents become first-class citizens of the OS, and ambience becomes the default way of using a computer. Evidence is early and mostly in preview. Windows is testing agentic features with Copilot Actions, MCP connectors, and separate agent accounts for Windows Insiders. Android's AppFunctions let apps expose tools that Gemini can call. Apple's App Intents let apps hand their actions to Siri and Apple Intelligence. Signal to watch: these leaving preview, and permission systems designed for agents rather than apps.
Through all three, a second economy grows underneath: agents paying other software to get things done.
Apps used to be where you went to use software. Now they'll be where the agent sends you when words aren't enough, and the best new customers won't be people at all.