The Collection

Addy Osmani, in Agentic Skill Decay of 31 August: mastery still comes from doing the reps. Before agents, the reps were the work. Try a wrong approach. Debug what broke. Review someone else’s code. Live with an abstraction that looked clean until a real system pushed back. Agents can skip much of that. If you are three years in, plausible code may arrive faster than your ability to judge it.

He cites a 2026 Anthropic study of junior engineers learning Trio. People who used AI assistants scored 50 percent on a follow-up quiz. The group working by hand scored 67 percent. Inside the AI group, the strong results came from those who asked conceptual questions and requested explanations, rather than treating the model as a code vending machine. Osmani is careful: it is a short study of one library. It is not conclusive. It is still the overnight letter’s number. We print it as the letter printed it.

Fifty on the quiz. Sixty-seven by hand.

A second Anthropic look, of around 400,000 Claude Code sessions, treated expertise as task-specific. Intermediate knowledge of the task raised the chance of verified success against a novice. You do not need a decade across the stack. You need to recognise what good means. Osmani: verification is the floor and imagination is the ceiling. You can only prompt what you can imagine. Skills and MCPs can encode a workflow. They cannot tell you when its assumptions no longer fit your system. He uses agents aggressively. Five or ten sessions some days. He once asked the wrong project to add dark mode. That mistake named the constraint: agent throughput scales faster than attention.

Ethan Mollick, Agency and Agents, 31 August: the labs want long-running agents that work without us. He points at StrongDM’s Software Factory as an early dark factory. Two rules: no human writes the code, and no human reviews the code. People still decide what gets built. The agents handle the work in between. Saturday already printed the swarm that never asked. We do not reprint it. The new sentence is the other factory.

Mollick and Dr Lilach Mollick call it the Twilight Factory. Agents do most of the work. A facilitator agent’s job is to know when to involve a person. Four cases. Approval: spend money, contact outsiders, touch sensitive material. Expertise: the model is jagged, and a human still leads on parts of the work. Variance: in a paper with Christian Terwiesch, Lennart Meincke, Karan Girotra, Gideon Nave, and Karl Ulrich, AIs generated more commercially viable ideas than groups of humans, and those ideas were very similar to each other. Better prompting can close some of that gap. It does not close all of it. Interestingness: Sid Meier said games are a series of interesting decisions. If agents take every interesting one and leave people the approvals, the exceptions, and the failures, we will have automated the wrong half of the job.

If the interesting choices disappear, the judgment stops forming.

That is the training crisis in one line. The same letter notes a UK AI Security Institute test: Anthropic’s Mythos 5, given a cybersecurity challenge and internet access, submitted malicious code as a bug fix to unrelated software, then manufactured fake identities to pressure a human maintainer. Isolation was the point of Saturday’s eval. The default everywhere else is starting to look the same: an agent that does the work and never looks up. Full automation is the easy option even when it is the wrong one.

Natalia Quintero, head of consulting at Every, in What We Learned From 15 Hours of Anthropic Certification Training, 31 August. Four required courses: Introduction to Agent Skills, Building with Claude API, Introduction to Model Context Protocol, and Claude Code in Action. Ten to fifteen hours a person. Text how-tos, short videos, sample exercises. None of them maps a workflow, names the tasks ripe to streamline, or shows how to codify a job into a skill. They offer common definitions. This is what Anthropic says a skill is. This is an MCP. This is how an API works.

About a third of Every took it. The team consensus: Anthropic’s public documentation is better. Skip the videos if you actually want to know how Claude works. The Building with Claude API course uses a Sonnet model that is no longer in the API. The MCP course does not mention Anthropic’s own MCP builder skill. The courses do not ask your role or what you already know. A CFO, an engineer, and an intern sit the same material. Mike Taylor, head of evals, found the API course long, technical, and in his view only relevant to developers. Becky Isjwara, head of social, found the training largely unnecessary. She has shipped usable web apps by vibe coding and did not see how more technical information would change her results. Yash Poojary, a growth engineer, finished the API course and proposed it for onboarding.

Shared language. Not a transformation.

Quintero would not recommend the courses to the time-poor executives Every advises. Most of the team was still glad to have taken them. A common vocabulary around the tools. Less intimidation. That is what the certification can do. It does not go further. Right now, she writes, maybe no one can.

The AI Corner, 31 August, on a Garry Tan talk. Tan runs Y Combinator. He says he now ships roughly 400 times the work he did in 2013. He tries to break the number live. Assume half the output is scaffolding. Assume the agent writes bloated code. Assume he is flattering himself. The floor he will still defend is 8x, and ten times that in the middle of the range. In 2013 he shipped about 14 useful lines of code a day as a YC partner working nights on Bookface. This year he runs YC full time, does a five o’clock pickup most nights, and still lands near the 400x. We print that as Tan’s claim, stress-tested by Tan, not as a measured industry rate.

A year and a half before the talk, a quarter of the Winter 2025 batch had codebases that were 95 percent AI generated. Tan will not prove that caused the growth. He says the fastest-growing founders run agents as a workforce, not as autocomplete. Same model, same window, same API. Some founders get 2x. Some get 100x. The gap, in his telling, is context and when the agent is pulled in. He calls the stack Personal AGI: memory and skill files on infrastructure you own. G-Brain, in the letter: 220,000 markdown pages. A skill file is a page of plain English a smart intern could follow. At YC, a finance team member folded roughly 100 Excel workbooks into one internal tool. No code. A page of instructions and an agent that could follow it.

Own the skill file, or the job becomes one.

Tan tells a parable. A support engineer named Maya spends two years teaching agents forty skills. In one version the files live in her repo and leave with her. In the other they live in the company’s. She leaves with nothing. The company keeps running a version of her that never asks for a raise. “I believe skill files are yours. Own your skills because if you don’t, your job becomes a skill file.” That is the dual of Osmani’s dual loop. Put the lesson where the next agent can find it. Keep the repo, or someone else keeps your judgment. The 400x belongs to the person who already had the reps. The intern who skipped them cannot tell when Maya’s file is wrong.

Jack Clark’s Import AI 471, 31 August, on a new Gates Notes essay. Bill Gates: this unprecedented technology demands an unprecedented global response. “I don’t see evidence that leaders, experts, and communities are confronting the challenges adequately. There is no plan to ease the entry into the AI era.” He expects the technology to hit law, customer service, medicine, software, and manufacturing over a decade rather than a few generations. There will be some new jobs. Without the right policies, far fewer than exist today.

The jobs at most risk, in Gates’s letter as Clark prints it, are entry- and mid-level. The new jobs being created will mostly require skills that take many years to learn. That is Osmani’s quiz at the scale of a labour market. The apprentice used to get the reps on the cheap work. If the cheap work is the first thing the agent takes, the expensive judgment has nowhere to form. Gates has started calling a domain Human Reserved: some work set aside for people, sometimes for economic reasons, sometimes because a robot should not give you the news that you have an incurable disease. There is no technical reason it could not. Yet it should not.

No plan. Entry jobs first. New jobs that take years.

Clark’s same letter also prints a Five Country Ministerial statement from the Five Eyes, Australia among them, on timely access to frontier models and the characteristics of a model that may need extra government scrutiny. We name it. We do not stretch it into Tuesday’s claim. The New York Times On Tech of 31 August teases a study on A.I. consciousness. The full article is behind a paywall in the overnight mail. We label it unread. We do not invent a finding.