My First Week With an AI Coworker

Context Window

My First Week With an AI Coworker

Plus: Why we’re skipping GPT-6.1 Terra for Gemini 4 Argon, and a prompt that keeps track of what you owe people

A week into my job at Every, one of my most helpful coworkers is an AI agent. It finds documents, catches details I’d have otherwise missed, and makes memes of our head of platform. It also gets things wrong with total confidence. Today, what I’ve learned so far about when to trust it.—Grant Harvey

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

---

What’s happening, and what matters

• Skip: GPT-6.1 Terra. OpenAI’s new model launched Monday, priced at half the cost of GPT-6.1 Sol and pitched for computer use (it’s fast at simple tasks like installing Google Chrome, but fails on more complex tasks like setting up this article in Every’s content management system) OpenAI also says it’s “significantly better at writing.” Well, is it? “We ran it but didn’t see enough of a difference to [run a] Vibe Check,” Dan Shipper says. I tested it on a basic writing task: drafting Thursday’s Context Window. The first version was full of AI tells; on the second and third tries, Terra took my notes well, editing its drafts better than previous GPT models. That makes Terra a capable iterative writer that still needs a lot of human steering. (No one here has tested its computer use yet, so maybe try it for yourself if you delegate desktop work to an agent.)
• Try instead, for editing: Gemini 4 Argon. Dan praises its ability to organize complex topics and find useful metaphors. He keeps leaving ChatGPT to use it, despite Google’s dated app: “When a model is good enough to force you to use a bad harness, it tells you something,” he said in our Vibe Check. Dan’s advice to the team this week: “highly highly recommend Argon.” But use it as an editor, not a ghostwriter: Its own drafts read at a dense 10th-grade level in our tests.
• Try: Have your Dot text you. Dots, OpenAI’s always-on personal agents, can now send text messages to your phone. Send one a task by text, photo, or voice note, and it texts you back when something needs your attention. (Dots are rolling out gradually to eligible ChatGPT Pro, Business Premium, and Enterprise users, with availability varying by region and workspace settings.) OpenAI’s Nick Baumann routes his Dot’s alerts by urgency: ChatGPT for regular updates, Slack for work, and a text for anything urgent. His setup line: “For any urgent matters, always prefer to send me a text via iMessage or RCS.”
• Important: It doesn’t matter whose idea it was, as long as you know why. In a new study, Anthropic researchers instructed Claude Code to favor randomly chosen approaches to a data analysis task, without telling the 50 analysts using it. The analysts chose Claude’s preferred approach 86 percent of the time. When explaining their decisions afterward, 77 percent repeated Claude’s arguments, but only 15 percent acknowledged that those arguments came from AI. Is that a problem? Dan doesn’t think so: “‘Mine’ is about whether I know why I made the call—what it means—not necessarily whether I originated the idea.” So before you sign off on your agent’s work, make sure you can explain each call it made in your own words and why it makes sense for you.

---

Signal

My first week with an AI coworker

What happened: I’ve spent my first week at Every working with an agent that’s basically another coworker. The Every Agent, the AI teammate we built, lives in our Slack, and you can add it to yours. When editor in chief Kate Lee told me Wednesday’s issue needed this week’s Thesis Statements, I asked the Every Agent where to find them. Less than a minute later, it had linked the doc, listed all nine contributors, warned me that one had changed his entry at the last minute without tracked edits, and told me to check that the web pages were live before publishing. Questions like that are one of the agent’s most common jobs, says Marcus Moretti, who leads the Every Agent: “I ask it dumb questions all the time.”

It’s also our office meme machine. Head of platform Willie Williams recently appeared in an Anthropic video, and a frame of him mid-interview got flagged by commenters comparing him to Steve from Stranger Things and a curly-haired Blake Shelton. Naturally, this became a Willie Meme skill: Describe a situation, and the agent captions that frame in the style of a subtitle. On Tuesday, after someone online bragged about getting past our paywall to read a post that was free, Dan asked for “a willie meme about this situation.” The agent delivered:

Dan: “Willie, you’ve been automated.” But the agent is not always right, and it doesn’t sound any less sure when it’s wrong. Last weekend, the agent told a teammate that their August time off was still waiting on approval when the request had been approved and logged months earlier. It had trusted an outdated record instead of the shared tracker. The slip went to the channel where the agent reports its own bugs. There, the agent team works through them with it: It can read its own code, explain what went wrong, and open a fix.

Why it matters: For me, the tricky part is triage: when to call on the agent, when to take something out of Slack for deeper work, and when to go to my coworkers themselves. Agents try to succeed at the task in front of them. Head of evals Mike Taylor put it this way: “LLMs are trained to win the game, humans are trained to play in a way that gets you invited to the next one.” So I start by asking what each job needs most: the team’s context, which the agent already has; a long stretch of focused work, which suits a chat like Claude; or judgment someone will stand behind, which only a person can give.

Marcus says the agent team has put “a huge amount of effort into its ability to index and organize and keep track of all the context it gets,” from our Slack to dozens of connected apps. It remembers what you and the team tell it, and skills, “playbooks teammates write and share,” turn one person’s process into everyone’s.

What it means: My working rule after a week: Use the agent for anything that touches the team—finding a document someone shared last month, chasing a thread you forgot to answer, posting a recap to a channel, polling five people at once. When you want to make something with that context, like drafting this newsletter, use a chat. And ask a human coworker for judgment, sign-off, or anything only they know.

Try this: Add the Every Agent to Slack, connect your team’s tools, and let it suggest an automation to help your team—setup takes about five minutes.

Then try these, adapted from what mine set up for me this week:

• “What can you do for me that I haven’t tried yet?” It reads back through what you’ve asked it. Mine found my biggest gap in seconds: “You have no scheduled tasks, so all your Context Window prep is still done by hand.”
• Ask the dumb question first. Before you interrupt a teammate, ask the agent. It has read more of your Slack and docs than any new hire, so you get an answer in seconds and save your coworkers for the questions only they can answer.
• Tell it what good looks like, not every step. A clear target lets the agent handle cases you didn’t script, and gives you something to correct it against. It’s how the agent team builds the agent itself: Define the target, watch how it performs, and tune from there.
• Ask where it checked. For anything you’ll act on: “What’s your source, and when was it last updated?”

---

Skill share

Keep a running tally of what you owe people

In her new 🔏guide to Dots, staff writer Laura Entis rounds up the ways the Every team puts them to work. Her own use is the simplest: keeping track of what she owes people. The requests she’s most likely to drop are the ones she agrees to in passing, like a quick “sure” in a Slack thread or an email. So her Dot keeps a running list, adding each request as it comes in and periodically messaging her about upcoming deadlines, feedback she promised, and decisions waiting on her.

Steal this prompt:

Go through [Slack channels / my email] and find anything I need to respond to or address: upcoming deadlines, unanswered questions or requests, or decisions that affect [specific projects]. For each, send me a link to the original message, a sentence on why it's important, and suggested next steps.

It works with any always-on agent (like the Every Agent), not just Dots. Laura’s guide has more workflows from the team—from a Dot that caught a $5,700 error in a retreat budget without being asked to one that runs multistep projects from a Slack thread.

---

Jagged frontier

Beyond docs and decks

A big part of a manager’s job is communication, usually in pursuit of alignment. And the biggest enemy of alignment is misunderstanding. Hence all the time we spend writing documents, making presentations, and explaining what the hell we meant by them.

In my last piece, I wrote about using feeds to understand what’s happening around you. But once you have that information, you still need to communicate it to other people.

There’s a familiar argument here about using documents versus PowerPoint to get that message out. The document people look askance at the slide people, and vice versa. Both have a point. A document gives you room to develop an argument and include the details. A presentation lets you move through the material visually, with someone explaining it along the way. There are tradeoffs around how much information you can convey, how quickly, and how much work you’re asking the recipient to do.

AI gives us more options.

Previously, choosing a format also meant choosing something you had to know how to make. A video meant learning editing; an interactive webpage meant assigning yourself a small engineering project to explain the actual engineering project. AI makes these options more practical because it does the building, even across multiple formats. AI allows you to use a narrated video to explain the reasoning behind static slides and an interactive staffing plan to let people toggle allocations and explore projected timelines.

Recently, Mike had AI interview him about his writing-skill experiment and turn it into a mini-documentary so anyone can hear the story behind the work. That’s a lively, scalable way to communicate ideas that might otherwise live in a lengthy document or require him to walk people through his process one by one.

Mike Taylor's mini-documentary on his writing skill, built with Claude.

The easiest way to start building your communication of choice is with existing assets. Pick a proposal, spreadsheet, or update from your feed and tell the AI who it’s for and what they need to understand. Then choose a format and give it a reference: a document template, a webpage, or a video in a style you like. Ask it to turn your material into that format, keeping the reasoning and assumptions intact. For a staffing plan, specify which allocations people should be able to change and how your spreadsheet calculates the timelines. You don’t need to invent the content and the format simultaneously.

Then review it yourself, share it with a small group, and iterate. That part hasn’t particularly changed.

For me, the best indicator of clarity is the quality of the questions I get back from the AI. If I’m describing an architecture, I want to get past questions caused by a confusing explanation and into questions about whether the design makes sense. A better format can help us spend our time on the tradeoffs that actually need discussing.

You want to spend less time explaining what you mean and more time discussing whether it’s the right thing to do.—Willie Williams

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. Every subscribers can RSVP for the after-party that evening for free.

---

Log on

Upcoming events

• All Access office hours: Friday, October 30, 12–1 p.m. ET, on Zoom: Host Natalia Quintero and the Every team share what we’ve been building and what’s coming up, then take All Access members’ questions about stuck agents, drifting workflows, and tools they’re unsure about adopting. All Access membership required. Register.
• Thesis: 2027: Our inaugural conference is one week from today, on November 5. Every subscribers can RSVP for the after-party that evening for free.

Past events

• Context Engineering With DSPy: Head of evals 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.