Apps, Agents, and Aggregation… | Stratechery Summary

1. Key Themes

Agents shift the scarce resource from discovery to inspiration

The abundance of apps and websites solved discovery (Aggregation Theory's original insight), but agents make doing things abundant too — leaving volition/inspiration as the new scarce resource, and whoever captures it becomes the ultimate gatekeeper.

"What is happening with agents is that the ability to do stuff is becoming abundant; what is scarce is volition. The problem to be solved is not discovery, but rather inspiration; the companies who solve inspiration will gain power over every entity that has things that need to be done."

Apps and enterprise UI were always just means to an end, not the end itself

Thompson argues apps existed only to expose functionality, and once agents can act directly, the interface layer disappears entirely — including in the enterprise, where inscrutable software UIs were merely lock-in mechanisms.

"no one wants to use apps because they are apps, apps were just the vehicle for people to accomplish something or entertain themselves; once people realize they can get straight to the job to be done it will seem odd they did it any other way."

Agents require a personal computer, not just a chatbot

The critical unlock isn't a smarter model but giving every user an actual provisioned machine so their agent can act on their behalf — this is what made Meta's Muse launch significant.

"It follows, then, that to actually deliver an agent that is usable by people unable or unwilling to set up a computer for their agent, you have to give people not just an agent but also a computer for their agent."

The new prize is becoming both the platform and the Aggregator simultaneously

Muse and Copilot aren't just chatbots or OSes — they're attempting to fuse platform dynamics (owning the interface) with Aggregator dynamics (owning demand), a combination Thompson calls the biggest prize in tech.

"This is the new prize in technology, and it is the ultimate one: not just a platform like Windows, or an Aggregator like Facebook. What Muse and Copilot are making a bid to be is both."

Incumbent distribution will determine early agent winners

Because agents get stickier with more context/access and most users will settle on a single agent, the companies with pre-existing reach and messaging infrastructure have a structural head start.

"Meta reaches nearly every person on earth; Microsoft reaches nearly every employee. The future may be distributed more rapidly than you expect."

2. Contrarian Perspectives

Generative, disposable, single-use apps beat scaled, shareable software

Conventional software wisdom prizes reusability and shared codebases; Thompson argues the future is apps built for an audience of one, thrown away after use, which inverts the entire logic of app development and distribution.

"the agent of my choosing can create any app that I want on command, even if that app is only for me... It's effectively disposable, because it's infinite."

Enterprise incumbents charging premiums to be plugins are cash grabs, not strategy

Rather than seeing Salesforce's move to charge extra to be integrated into Claude as savvy platform positioning, Thompson frames it as a doomed rent-extraction attempt on infrastructure destined to be bypassed by computer use.

"Salesforce is going to try and charge a three-digit premium to make their flagship product a Claude plugin; I salute the audacity of the cash grab in the face of computer use that will ensure the Salesforce UI is never interacted with by a human again."

Models themselves are not the differentiator — distribution and accumulated context are

Against the popular narrative that model quality (benchmarks, capability races) determines AI winners, Thompson argues substitutability of models means the real moat is context and access accumulated by the agent layer.

"Models are relatively substitutable; agents, however, operate better the more context they have about you, and the more access they have to things like your logins and files. That increases their stickiness."

3. Companies Identified

Meta — Social media/AI company that launched Muse and provisions users with virtual machines via Meta Connect devices. Mentioned as the first mover giving ordinary users "a computer for AI," and reframing its glasses/AR hardware strategy entirely.

"Meta's devices are not AR or VR devices, or even AI glasses: they are Muse delivery mechanisms... the general purpose computer I was talking about exists: Meta gave it to every user for Muse to use."

Microsoft — Enterprise software giant relaunching Copilot as a "super app" combining chat, coding, and agents (Autopilot). Mentioned as pursuing the same "OS for work" strategy it's held since Teams, now instantiated via agents with enterprise-grade governance.

"Autopilot lives in your tenant with its own identity, memory, computer, and workspace, and it's built on Microsoft IQ so it understands how your organization actually works."

OpenAI — Referenced via ChatGPT's evolution from chatbot to computer-operator, foundational to Thompson's agent thesis.

"And now ChatGPT has a computer of its own."

Salesforce — Enterprise CRM company attempting to charge a premium to be integrated as a Claude plugin at Dreamforce. Cited as a cautionary case of incumbents trying to monetize their soon-to-be-obsolete UI layer.

"Salesforce is going to try and charge a three-digit premium to make their flagship product a Claude plugin."

Facebook (historical Facebook Home / WhatsApp acquisition) — Referenced as a historical case study of a failed people-centric interface strategy versus the successful jobs-to-be-done messaging bet (WhatsApp).

"Apps aren't the center of the world," he says. "People are."

Apple — Referenced via the iPhone launch as the framing device for the entire piece, illustrating how revolutionary shifts are often misunderstood at launch.

"Today, Apple is going to reinvent the phone."

Wolfram|Alpha — Cited as an early ChatGPT plugin example demonstrating AI's use of deterministic computing tools.

"The fact this works so well is itself a testament to what Assistant AI's are, and are not... And now ChatGPT has a computer of its own."

4. People Identified

Steve Jobs — Apple co-founder; his iPhone launch keynote is used as the central analogy for how the market fails to grasp genuinely revolutionary product shifts.

"Are you getting it?"

Mark Zuckerberg — Meta CEO; quoted from 2013 defending Facebook Home's people-centric philosophy, later revisited as the seed of Muse's agent/computer strategy.

"Apps aren't the center of the world... People are."

Satya Nadella — Microsoft CEO; confirmed via X that Microsoft's agent strategy is a continuation of its long-standing "OS for work" ambition.

"Satya Nadella made clear on X that Microsoft's strategy was the same as it ever was: be the OS for work."

Jared Spataro — Microsoft VP; described Autopilot's enterprise-grade design and integration into existing Microsoft workflows.

"It shows up where people already work — Teams, Outlook, chats, channels, and documents — so you can @mention it like a colleague, with permissions, audit, and governance behind it."

Ben Thompson — Author; uses his own personal experience (Muse-built recipe app, declining app usage) as primary evidence for the thesis.

"the number of apps I look at are plummeting."

5. Operating Insights

Build for disposability, not scale. The economics of software creation are inverting — instead of building durable, shareable products, operators should consider ephemeral, single-user tools generated on demand, since agent-built software is now essentially free to produce.

"It's effectively disposable, because it's infinite."

Distribution + context, not model quality, is the moat to build for. Since models are commoditizing, founders and operators building AI products should prioritize accumulating user context and system access (logins, files, environment) over chasing marginal model improvements.

"Models are relatively substitutable; agents, however, operate better the more context they have about you, and the more access they have to things like your logins and files."

Give agents a "computer," not just a prompt box. Any AI product aimed at mainstream (non-technical) users needs to abstract away setup complexity by provisioning actual compute environments (VMs) behind the scenes — this is the tactical difference between a toy chatbot and a usable agent product.

"to actually deliver an agent that is usable by people unable or unwilling to set up a computer for their agent, you have to give people not just an agent but also a computer for their agent."

6. Overlooked Insights

Rate-limiting by legacy platforms (e.g., Instagram) is already an active friction point against agents, hinting at a coming conflict between platforms that want to protect their walled gardens and agents attempting to act on users' behalf across services — a potential regulatory/business battleground not fully explored in the piece.

"it's working on it, subject to Instagram rate-limiting."

Computer-use capability quietly became viable for any GUI app, not just command-line tools, starting with a specific model version — suggesting a concrete technical inflection point most readers may have missed amid the broader agent narrative.

"every app with a command-line has been usable by agents for a long time; starting with GPT-5.6 Sol, every app with a user interface was usable too, albeit slowly. Astra made it fast."