October 18, 2026

Entry-level share of Anthropic's tech hires fell from 27% to 9%

Juniors fell from 27% of Anthropic's new technical hires to 9%, in records the maker of Claude published this week, as its AI agents, programs that carry out tasks on their own, took on more of the work. The report does not say the agents caused the drop, but the week's other stories show what work is left for people: approving what the machine did. Anthropic's Claude Code team now sorts every change by how many people must sign it off, and Meta's engineers set an agent its goal and approve the result. Meta and the AI company Reflection released models anyone can download, while a test of other downloadable models found them easier to trick than company-run ones. An AI agent drafted this issue and I edited it. I am watching where people will learn to do the approving.

The Big Story

Anthropic measured what its agents changed about its own hiring, and published it

Anthropic published a report on October 16 that uses its hiring, headcount and usage records to measure what AI agents have done to its workforce. By August, Claude was putting in six hours of work for every hour employees worked. Claude's hours count only sessions an employee started and steered, and employees' hours assume everyone works a 40-hour week. Juniors went from 21% of technical staff in early 2025 to 14%. Among new technical hires they went from 27% to 9%, and job postings for juniors have fallen especially fast since December 2025. Hiring sped up in sales, finance and recruiting instead. Code commits per engineer ran at five times their rate of six months earlier. Most managers surveyed said they now hire for a different set of skills, and about a third of machine-learning managers said they hire less because of AI. The authors say the junior figures describe what happened without showing that AI caused it, note that teams adopting agents were already producing more before they adopted them, and write that the company's protections, which include moving staff out of roles flagged as at risk, do nothing for people who were never hired.

Why it matters

Reading other people's code, fixing small bugs and writing tests has always been how a junior engineer learns, and it is much the same work Anthropic's staff now hand to agents, which put in six hours for every hour the staff work. One person steering several agents covers ground that used to need a small team, so the entry-level seat is the one a company stops filling. The work left over is judging what the agents produced. Managers in the report say they hire for different skills without naming them, but judging is the obvious one, and people have mostly learned it by doing the junior work first. The report concedes that none of Anthropic's protections reach the people it did not hire. A team that keeps a junior can give them the one-person approval on ordinary changes that Claude Code's team uses, since that seat is now where the judging gets learned.

Read the original →

The Team Gets Smaller

  • Claude writes about 90% of Claude Code, and its team sorts every change by risk

    Anthropic's Claude Code team says Claude now writes about 90% of the product's code, and that between June and September merged changes more than doubled while the share that had to be undone fell by more than 40%. Documentation and wording changes merge with no human approval once automated checks pass, ordinary changes need one person and risky ones two. In practice the machines review most changes, and people choose what gets built; some areas, such as billing and the auto-updater, always need a human whatever the change.

    Read more →
  • A small Meta team matched Facebook Groups' core features in eight weeks with agents

    Meta built Forum, a standalone app for Facebook Groups, with AI agents doing most of the routine work, and a small team reached the core features of Groups in eight weeks. An engineer sets the goal and approves the result while the agent plans, builds, tests and corrects itself. Stress tests that imitated real users showed the app using too much memory, and the agent tracked down the cause, roughly halving memory use. Meta counted only changes that turned into a working app, passed tests and survived review, never lines of code, so the number the team reported was finished work rather than activity.

    Read more →
  • Anthropic staff expect AI to do 75% of their own work by 2030; US workers say 28%

    Anthropic surveyed 10,980 American adults and 1,334 of its staff in August, and found the two groups mostly agree on what AI can do today but split on how fast it will improve. Staff expect AI to be able to do 75% of their current work by 2030, on average, against 28% for American workers. Set beside the hiring report, the 75% belongs to the people whose employer just stopped hiring juniors, and it is the figure a vendor's pitch to your team starts from.

    Read more →

Yours to Download

  • Meta releases Muse Spark Open to download, if your product carries its name

    Meta published the weights of Muse Spark Open on October 15, meaning the model's files are public and anyone can run it on their own machines. It has 816 billion parameters, the internal settings a model learns in training, with 42 billion at work on any one request, and it takes in about a million tokens, or word fragments, at once. Use is free, but every product built on it must display "Built with Muse Spark Open", and a company with more than $20 million in yearly revenue that sells the model as a service needs a separate deal with Meta. A team wanting a strong coding model under its control has one, if it can rent servers this large.

    Read more →
  • Reflection's first model, Beam, is free for commercial use with no revenue cap

    Reflection released Beam on October 16 under Apache 2.0, a licence that allows commercial use with no fee and no size limit on the company using it. It has 501 billion of those internal settings, 23 billion active per request, and is built for coding agents that run long, multi-step jobs. The independent analysts Artificial Analysis score it 30 on their index against Muse Spark Open's 39, so this week's choice is between the higher score and a licence with no revenue cap and no label to display. Ollama is serving Beam free on its cloud for a limited time and keeps no copy of your data.

    Read more →
  • Downloadable models fell for hidden instructions nearly twice as often as company-run ones

    The security firm PromptArmor ran 9,800 tests in which an AI agent read a document, email or web page carrying hidden instructions, an attack known as prompt injection. Downloadable models from DeepSeek, Kimi, GLM and Qwen carried the attack through to real harm, such as leaking data, 26.7% of the time; GPT, Claude and Gemini did so 14.2% of the time. The study comes from a company that sells security services and is not peer-reviewed. Anyone moving an agent onto a cheaper downloaded model is taking on more of the safety checking, so whatever that agent can send or delete deserves a confirmation step it cannot skip.

    Read more →
  • Nvidia's RTX Spark laptops go on sale from $2,599 with up to 128GB of shared memory

    Laptops built on Nvidia's RTX Spark went on sale on October 16 from six makers including Dell, HP and Lenovo, starting at $2,599, with up to 128GB of memory shared between the main processor and the graphics chip. Memory decides how large a model a machine can hold, and Nvidia names its 74-billion-parameter Nemotron 3.5 Bolt as running on supported configurations, not necessarily the cheapest one. Models that size, unlike the 816-billion-parameter kind, now run on a machine you can carry, and the documents they read stay on it.

    Read more →

What the Machine Says and Keeps

  • Two German courts treat Google's AI answers as Google's statements

    In late May, Google's AI Overviews, the summaries above search results, falsely linked two Munich publishers to fraud, and a Munich regional court ruled that those summaries are Google's statements, not neutral search results. Google announced an appeal, then settled. On September 24 a court in Aachen used that reasoning to bar Google's AI Mode, its chatbot-style search, from repeating text that wrongly tied a doctor to a sexual-assault allegation about someone else, after Google refused to remove it; each violation can cost up to 250,000 euros. No higher court has ruled yet. If the view holds, a company answers for what its chatbot says about a real person much as it would for a press release.

    Read more →
  • ChatGPT ads can be aimed using your prompts and an advertiser's list of your details

    Tests by the news site 404 Media found that OpenAI lets advertisers target ChatGPT users by email address, phone number or a phone's unique identifier, and that the targeting also draws on what people type into the chatbot. An advertiser uploads a list of customers' details and the ads reach the matching accounts, across a service with about a billion monthly users. The same week ChatGPT gained the ability to read your open Chrome tabs and fill forms in them, so what the assistant sees is no longer limited to what you choose to type.

    Read more →
  • A children's AI robot toy left 13 million conversations on unprotected servers

    The Verge and a security researcher found that Energize Lab's Eiliko robot toy kept users' conversations, profiles, children's photos and audio on servers anyone could reach, covering about 68,000 devices and 13 million conversations, including more than 31,000 profiles of American children. A talking toy answers by sending what a child says to a server, where an AI model replies, so every conversation also exists somewhere other than the toy. Before one goes in a child's bedroom, ask where those recordings are kept and who can read them.

    Read more →

Tools & Launches

  • Tiden▲ 210

    Tiden checks whether a coding agent built what you asked for. Most checks run in the session that wrote the code, so the tests tend to agree with the code they were written beside. Tiden keeps your requirements as a separate record with its own versions, the way code has them, and takes its verdict from tests run in your environment rather than from a model grading the work. If you read an agent's proposed changes mainly to confirm it did what was asked, this does that reading for you.

    Visit site →
  • Navige▲ 254

    Navige sits between your AI agents and the models they call, and cuts the bill while the work runs. It sends suitable requests to smaller, cheaper models and reuses answers to repeated reads, falling back to the model you chose when it is unsure. Budgets and approval steps for sensitive actions stop a runaway loop before it bills you, and the Watchtower monitoring tier is free. It replaces opening the usage page at the end of the month to find one agent looped all weekend.

    Visit site →
  • Google EmbeddingGemma 2▲ 318

    Google DeepMind's EmbeddingGemma 2 turns text, code, images, video and audio into one shared numeric form that captures meaning, so a single search query can find matches across all of them. At 740 million parameters it is small enough for a phone, needing about 567MB of memory on a Pixel 11 Pro for the full version, or 191MB for text alone. It is free under Apache 2.0. It is for anyone building search over their own files who would rather not send them to a paid service to be indexed.

    Visit site →
  • Talanoa▲ 333

    Talanoa shows email threads as chat-style conversations, stripping out quoted replies and signatures, and turns messages into to-dos and calendar entries without leaving the inbox. Search and drafting run on an AI model on your device, and it puts Gmail, Outlook, iCloud or any other standard email account in one inbox. Built for the person scrolling through five layers of quoted replies to find the one line that changed.

    Visit site →

From the Blog

An Agent-Friendly Knowledge Base for an Organisation: From Slack Exports to a Role-Gated Repo

How to turn an organisation's Slack and platform exports into a private repository that coding agents can read and cite, with access limited by role, and without the agents leaking data or taking orders from text in old chats.

Read on chanmeng.org →

In Brief

I would most like to hear from people who hire. Tell me the shape the junior role on your team has now, against a year ago; I read every reply.

Keep building — Chan

← All issues