Apps, Agents, and Aggregation (Stratechery Article, 2026-09-28)
Why this is in the vault
Thompson extends Aggregation Theory to agents: abundant discovery gave way to abundant doing, so the scarce resource is volition and the power sits with whoever supplies the one agent a person or company uses. This is the cleanest statement yet of the frame that ties his recent Muse, headless-SaaS, and memory pieces together. Full body rendered from email (no paywall truncation).
Mapping against Ray Data Co
Mapping strength: strong. The concrete anchor is Ray's own operating posture. Ray drives Gmail, Calendar, Notion and Slack through MCP tool calls and never touches their UIs, and Thompson's claim that apps become "implementation details" behind the agent is that same pattern described from the market side. His Muse recipe-box anecdote (agent builds a disposable custom app in 5 minutes while he walks the dog) also mirrors how RDCO builds one-off HQ decision pages and artifacts on demand instead of adopting a product.
Two sharper implications. First, his stickiness argument (models are substitutable, agents accrue context, logins, and files, so most people keep one agent) cuts against any strategy of being a second or third agent in someone's life; it supports keeping the RDCO/phData angle on serving inside an incumbent agent's distribution (Claude in the enterprise) rather than competing for the front door, which is consistent with [[project_credibility_for_phdata_sales]] and the "agents in production" domain focus. Second, his "computer per agent" point (Meta gives every user a VM; Microsoft's Autopilot has its own tenant identity, memory, and workspace) is a useful outside validation of Ray's Mac Mini always-on setup as the right primitive, and of the write-things-down memory layer as the durable part.
Contrarian check: Thompson concludes the winners are the two distributors (Meta for people, Microsoft for employees), which is a bearish read for small independent agent products; treat it as a reason to pick niche and integration bets, not as evidence Ray should build a general agent.
The core argument
- Recalls Jobs' "are you getting it" iPhone launch line and applies it to three claims: messaging/chat is how people talk to AI, AI will use computers rather than people, and pre-built write-once UI is dying.
- Callback to his 2013 "Apps, People, and Jobs to Be Done" and 2014 "Messaging" pieces: apps were vehicles for jobs to be done (151 apps then, 689 now), and most people never wanted to be app users.
- Agents are AI plus a computer. He ties this to his 2023 "ChatGPT Gets a Computer" and the write-things-down memory idea. Meta's real Muse move is a per-user VM (2 cores, 8GB RAM, 8GB storage), more than the Muse Spark model.
- Computer use (GPT-5.6 Sol made any UI usable, Astra made it fast) means agents can operate any app or site the user does not want to.
- Discovery shifts to inspiration: on the web, distribution was free and discovery was the scarce problem; with agents, doing is free and volition is scarce. Whoever solves inspiration gains power over every service, which becomes a supplier to the agent.
- Microsoft's Copilot "super app" (chat, code, and Autopilot, with a cloud computer, tenant identity, and governance) is presented as the enterprise twin of Muse. Both bid to be platform and Aggregator at once.
- Enterprise SaaS is exposed: he mocks Salesforce's premium to be a Claude plugin, given computer use makes its UI moot. Distribution advantage (Meta for people, Microsoft for employees) is why those two lead.
Sponsorship and bias
No sponsor block. Newsletter is Thompson's own paid product; footer is subscription boilerplate only. Bias to note: he is an admitted heavy agent user and Muse enthusiast writing in a period when he is predicting adoption, so the "it is obvious if you use agents" framing is advocacy, not evidence of mass adoption. He also flags that many people are not yet using agents.
Related
- [[2026-09-22-stratechery-amazon-blocks-muse-aggregator-moat]]
- [[2026-09-16-stratechery-salesforce-ai-force-agents-as-ui-headless]]
- [[2026-09-09-stratechery-ben-thompson-openai-math-muse-agent]]
- [[2026-09-08-stratechery-write-things-down-harness-memory]]
- [[2026-09-25-stratechery-aggregator-wars-gm-carplay-digest]]
- [[project_credibility_for_phdata_sales]]
- [[project_l5_north_star_strategic_direction]]