Agents Choose What You Read Now

Context Window

Agents Choose What You Read Now

Plus: Every automates our memes, an Emmy-winning TV writer’s AI writers’ room, and what Anthropic’s buy of a one-person biotech means

Hello, and happy Sunday. It was a week of agents at Every: Evan Armstrong works out whether Anthropic is worth $2 trillion, head of evals Mike Taylor tested which titles get an agent to recommend your work, and new staff writer Grant Harvey spent his first week with the Every Agent. We’re off next week for Think Week, when our team flies in from around the world to spend time together in person, play, and explore the future. We’ll send some of the best of what we’ve published, plus a new podcast. And on Thursday, November 5, we’re hosting our first conference, Thesis: 2027. Anyone with a free Every subscription can watch all 27 sessions, and viewers get a shot at $20,000 in AI credits and benefits from partners including Notion, Anthropic, and OpenAI. Register for the livestream.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

Overheard at Every

What the team couldn’t stop talking about this week

Turn your effort dial down. Kate’s reaction to a new paper from Johns Hopkins and Together AI: “Effort level!!” The researchers tested the reasoning-effort setting built into Claude Code, Codex, and ChatGPT. Maximum effort raised accuracy overall, but it often failed tasks that low effort had solved, and on easy tasks it cost up to 42 times as much. Their fix: Start at the lowest setting and raise it only after a checked failure.

The coding interview is dead, at least on one Anthropic team. Felix Rieseberg, an engineering manager at Anthropic, wrote this week that he stopped giving coding interviews because Claude now beats candidates at the small challenges they test. His bigger point is that newer engineers can now build a working prototype over a weekend instead of waiting years for permission. Engineer Lee Knowlton called it “another banger.” For now, Felix looks for small prototypes that show taste.

The Arc breakup continues. Head of social Becky Isjwara, a self-described “arc loyalist,” switched to Dia after it opened the FigJam from her call and grouped it with the Google Meet tab. Lee isn’t there yet: Dia is “like arc but uglier.” Customer support specialist Waqqas Mir moved to Brave instead: “cant go through another breakup with a browser.”—Grant Harvey

---

Protocols

From this week’s stories, 13 skills to try.

Get agents to cite your work

From Mike Taylor’s “How to Get AI Agents to Recommend Your Articles”

• Write the subtitle as the answer. Rewriting both the title and subtitle of Mike’s prompt engineering guide took it from two agent citations in 240 sessions to 240 of 240. The rules
• Bet on one post per topic. Agents tend to cite a single post from a site on each topic, so pick the one you want to win and make its subtitle the clearest answer. The rule
• Test your titles with agent-pass. Mike’s Claude Code skill reviews a post’s title and subtitle, suggests truthful rewrites, and tests them against real agents. The skill

Get more out of new models

From Grant Harvey’s “Every Model Has a Six-month Shelf Life Now”

• Brief AI on your taste before it vets your idea. Emmy-winning TV writer Fred Graver writes one doc about his inspirations, tastes, and ambitions and another about the project, then asks AI whether the idea really fits his voice. The method
• Give each model its own job. His “very unpleasant” Claude agent holds him to existing commitments, ChatGPT plays an eager film student who checks the basics, and NotebookLM’s podcast hosts supply encouragement. The cast
• Test each new model on one hard task. When a model ships, read the Vibe Check, then try it on your personal benchmark: one hard task you know well. The habit
• Ask AI which of your skills to sell first. Mike’s shoggoth strategy is to market yourself on one clear specialty and keep your generalist skills behind it. His prompt asks your agent to list your 10 strongest skills and pick the one you should lead with when people ask what you do. The prompt

Work alongside an AI coworker

From Grant Harvey’s “My First Week With an AI Coworker”

• Check your blind spots. Ask your team’s agent, “What can you do for me that I haven’t tried yet?” Grant’s agent found his biggest gap in seconds: He had no scheduled tasks. The prompts
• Check the agent’s sources. Before you act on an agent’s answer, ask, “What’s your source, and when was it last updated?” The prompts
• Keep a running tally of what you owe people. Staff writer Laura Entis’s Dot collects the requests she agrees to in passing and messages her about deadlines, promised feedback, and decisions waiting on her. The prompt
• Pick the format before you write the doc. Head of platform Willie Williams gives AI an existing proposal or spreadsheet, says who it’s for, and asks for it as a video, webpage, or interactive plan modeled on a reference he likes. The steps

Decide what to hand your coding agents

From Loren Stewart’s “What Karl Marx Would Say About My Coding Agents”

• Give each task a mode. Engineer Loren Stewart lets agents run on their own for routine work like his morning briefing, adds checkpoints before key decisions on systems he wants to understand, and codes by hand when he wants to learn something new. The modes
• Make the agent’s code teach you. For code he didn’t write, he asks an agent to write checks for the parts he wants to understand, then works through them in a debugger to see how the code works and what the checks missed. The method

---

Log on

Get hands-on with how Every uses AI at live camps, workshops, and meetups.

Upcoming event

• Thesis: 2027: Our inaugural conference is on Thursday, November 5. Watch all 27 sessions for free on the livestream. Register.

Past event

• Context Engineering With DSPy: Mike Taylor, author of the new O’Reilly book Context Engineering with DSPy, joined Drew Breunig and DSPy lead maintainer Isaac Miller to build a DSPy program you can measure and improve. Watch on YouTube.

---

Knowledge base

“Is Anthropic Worth $2 Trillion?” by Evan Armstrong/Napkin Math: Former Every writer Evan Armstrong reads Anthropic’s IPO filing and finds numbers unlike any public markets have seen. A $2 trillion price depends on the next six models being as good as the last six.

“Every Model Has a Six-month Shelf Life Now” by Grant Harvey/Context Window: Anthropic’s filing shows its revenue rides on its newest models, which makes switching to each new one a good habit for the rest of us, too. Plus: The shoggoth strategy for generalists.

“My First Week With an AI Coworker” by Grant Harvey/Context Window: The Every Agent finds our documents and makes memes of our head of platform, but it can also be wrong with total confidence. Here’s a working rule for when to ask the agent, a chat, or a coworker.

“How to Get AI Agents to Recommend Your Articles” by Mike Taylor/Also True for Humans: Head of evals Mike Taylor ran 2,484 sessions with Claude and ChatGPT to learn which titles agents cite, then turned the results into seven rules for writing for agents.

“What Karl Marx Would Say About My Coding Agents” by Loren Stewart/Source Code: Engineer Loren Stewart’s agents now write nearly all of his code. He uses Karl Marx’s theory of alienation to name what he lost and splits the work into three modes: automate, supervise, or code by hand.

🎧 🖥 “Can You Use AI to Make Good Art?” by Natalia Quintero/The Every Podcast: Emmy-winning TV writer Fred Graver tells Every’s head of enterprise AI adoption Natalia Quintero how he gives Claude and ChatGPT different jobs and why AI can’t supply the quirks that make characters work. 🎧 🖥 Listen on Spotify or Apple Podcasts, watch on X or YouTube, or read the transcript.

Read nine more predictions from frontier builders and thinkers in Thesis Statements, a collection of specific, contestable claims about the future of great human work with AI.

• Taste will be worth something only as long as humans are buying by Ben Cera, founder and CEO of Polsia
• We taught AI how to talk to us—now it’ll help us talk to each other by Yasmin Green, CEO of Jigsaw
• We’ll keep the “pointless” parts of our lives—because sometimes they’re the whole point by Senka Hadzimuratovic, head of communications at Lovable
• Baby is the prompt by Ashley Mayer, cofounder of Coalition Operators
• Software is art by Josh Miller, cofounder and CEO of The Browser Company
• Progress in the physical world will accelerate, looking more like the speed of software by Lauren Reeder, partner at Sequoia Capital
• AI will translate what we think for people who don’t think like us by Alexandra Samuel, technology writer
• Organizational value only scales as fast as your culture by Ryan Vauk, vice president of AI transformation at Google
• People will still find a way to be miserable by Visakan Veerasamy, writer and author

Our inaugural Thesis: 2027 conference is on November 5, 2026. Anyone with a free Every subscription can watch the livestream.

---

Alignment

Professor Oak, unmasked. After six years of anonymous X biotech commentary, “Professor Oak” revealed himself this week to be Cody Tranbarger, founder of Mithril Medicine. He did so by announcing that Anthropic had acquired his company, which is a beautifully spectacular way to introduce yourself to the world.

As Mithril’s sole founder and employee, Cody used Claude to help coordinate work across 11 outside laboratories. According to Anthropic, a potential treatment for a rare muscular dystrophy went from an idea to promising results in mice in less than 18 months. He now plans to start several more programs inside Anthropic to see how far Claude can extend his own expertise.

Anthropic is most likely going to turn that experience into better tools to sell through its relationships with pharmaceutical companies while applying the same improvements to its own drug programs.

For other founders reading the news, the question becomes how much company they need to build around their expertise. If Claude can reliably coordinate outside laboratories and take on more of the scientific work itself, a small team might be able to pursue several programs without the need to scale hiring.

Clinical trials will still take years, but expert drug hunters may be able to reach the point where a treatment attracts a buyer or further investment with far less time and money. Whether Claude can make that happen often enough and on compressed timelines will determine if these lean biotechs become a solid venture bet.—Ashwin Sharma

Thanks to Rebecca Ackermann for editorial support.

Stop explaining AI to your team and start showing them. The Every Agent does the work in Slack, where everyone can see it, so one good workflow quickly becomes the whole team’s.