Stratechery: the agent as the ultimate aggregator; Lenny's: Jev, a model that only decides

AI strategy: the agent as the ultimate aggregator

"Apps, Agents, and Aggregation" — Stratechery

1. The claim sits in Thompson's summary line, and the mechanism is his own aggregation theory. Agents are the ultimate aggregators, he writes, and providing one is tech's biggest prize. The theory runs the way it did for the web: distribution became free, content became abundant, so discovery turned into the scarce step, and the companies that solved discovery collected power over their suppliers. Agents make doing things abundant, which moves the scarce step again, this time to volition. Whichever company solves inspiration, in his argument, gains power over every entity that has something that needs doing, and apps and services shrink into implementation details behind the agent.
2. An agent runs a computer, and the author's own demonstration is an app he built in five minutes. Thompson's definition is an AI with access to a computer, which is why the part of Meta's Muse launch he fastened on was the virtual machine behind it rather than the model: two cores, 8GB of RAM and 8GB of storage, provisioned for every user in the United States. That machine is what hands an agent to people who will not set one up themselves. He asked Muse to organize the recipes he had saved on Instagram; the first attempt arrived as a PDF, he asked for something more usable, and within five minutes he had a working app, his 690th, built while he was out walking his dog. Today it is really a categorizer of Instagram videos, slowed by Instagram's rate limits, and that is the point he draws from it: a user interface generated for one person needs nobody else to be able to use it, and can be discarded once the task is done.
3. Microsoft and Meta are each bidding to be a platform and an aggregator at once, and both start with distribution. Microsoft unveiled its new Copilot super app on September 25, combining chat, a Code tab that lets anyone build an app, tracker or dashboard and share it with colleagues, and Autopilot, a long-running agent that keeps working while its owner sleeps. Autopilot lives inside the customer's own tenant with its own identity, memory and workspace, and colleagues can @mention it. 2 Satya Nadella called it "a new OS for work" 3, and Thompson reads the move as the next turn of the argument he made about Teams in 2020: own the interface, and require everyone else to integrate on your terms. Copilot does that with an agent and adds enterprise controls on top; Muse makes the same play for consumers. His reason for taking the contest seriously is stickiness. Models are replaceable between providers, while agents accumulate context, logins and files, so most people and most companies will end up with exactly one — and the two companies that reach almost every person and almost every employee hold the advantage.

Product and tooling: a model that only returns decisions

"Jev for beginners: how to use it and what to build" — Lenny's Newsletter (How I AI)

1. Jev is a decision model, and that single property is where the cost comes from. TypeSafe AI's Jev returns type-safe structured values — a category, a score, a probability — in place of generated text. It costs 4 cents per million input tokens and charges nothing for output, because there is no output to charge for. Vo puts its coverage at roughly 90% of what many software workflows actually need, and the question she asks before every build is now simply where a workflow only needs a decision. Classification, routing, ranking and filtering all qualify; writing documentation and designing an interface do not.
2. She ran it on five of her own projects, and the numbers carry the argument. Comparing 1,700 ChatPRD pull requests against one another, 17,000 pairs, cost 9 cents and about two minutes, and showed that close to 30% of the company's engineering work had gone into platform, security and infrastructure. Classifying the Claude Code and Codex sessions stored on her laptop showed engineering falling from nearly 100% of her AI use in January to under 40% by September, with agents and media publishing taking up the rest. ChatPRD's product-insights graph took 1,100 signals through roughly 200,000 operations for about $4. A dashboard built from 4,500 How I AI YouTube comments sorted them by sentiment and turned up 58 that contained episode ideas.
3. It sits inside a pipeline beside a frontier model, and it makes real-time loops practical. Vo's pattern is to let Jev classify, cluster, filter and route the large set, then send only the groups that matter to GPT-6 Astra for reasoning. Her demonstration of its speed is a voice app that turns a spoken phrase into a colour, matches a quote to the sentiment, and shows both at once; the quote service, not the model, turned out to be the slowest step in the loop.

Where to start