Everyone Gets an Agent. Almost No One Gets the Model.

Hello, and happy Sunday! OpenAI released GPT-5.6 Sol on Friday—and by U.S. government directive, access is limited to roughly 20 pre-approved companies while Washington works out how to release frontier models with advanced capabilities. Dan Shipper argued that the lockout hurts the people who most need these tools to keep up: “A world where advanced models are locked up only for use by the employees of AI giants and a select few companies is one where ambitious students, independent builders, and working professionals are denied the tools they need to learn, create, and compete to their fullest potential.”

That rationing is coming for the rest of us, too. Mike Taylor argues that token access is about to be allocated like capital—the biggest budgets going to whoever can prove the biggest returns, a dynamic that will only grow as models get more capable. From there, we got into what happens when people do get these tools in their hands: The agents built for engineers are coming for everyone else’s desk, with Codex crossing 5 million weekly active users and Anthropic’s Claude Tag landing in Slack. Our Compound engineering plugin can now run a coding agent unattended for hours, long enough to build a feature, write its tests, and open a pull request on its own. Nityesh Agarwal makes the case that Claude Code might be the only agent-builder you need, Katie Parrott hands her career review to Codex, and Dan Shipper asks Surge AI’s Edwin Chen what’s left for us once machines can do everything.—Kate Lee

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

Knowledge base

“Claude Code Is the OpenClaw Alternative You Already Have” by Nityesh Agarwal/Source Code: Senior applied AI engineer Nityesh Agarwal argues that anyone reaching for OpenClaw to build AI agents already has a more capable option: Claude Code, which Anthropic shipped over a year ago. Marketed as a coding tool, it went largely unrecognized as a general-purpose agent harness. He breaks down what that framing hid, and how the same tool became Claudie, Every’s always-on AI employee in Slack.

🔏 “Codex for Knowledge Work” by Katie Parrott/Guides: Our guide to using Codex for knowledge work—not just coding, but email, writing, research, and planning—is now in its second edition, barely a month after the first. We rewrote it to keep up with how fast Codex is moving: It maps how projects, threads, Goals, plugins, and Sites fit together; explains how Codex reaches your files and apps; and adds a 30-day plan to get you up to speed. One tip inside: Hand Codex the whole guide, tell it your role, and let it pick your first workflow.

“Codex for Everything and Everyone” by Laura Entis and Katie Parrott/Context Window: OpenAI is pushing Codex well beyond coding, with Sites and role-specific plugins for analysts and product managers, betting it becomes the place everyone gets agentic work done. Also: a Codex hack for YouTube thumbnails, Anthropic’s new Claude Tag Slack agent, and the AI tells now creeping into design.

“Token Tightening” by Laura Entis/Context Window: The era of measuring AI adoption by raw token consumption looks to be ending. Head of tech consulting Mike Taylor argues that compute will be allocated like a trading portfolio—the biggest budgets going to the few who can prove the biggest returns, with frontier access rationed accordingly. Also inside: head of growth Austin Tedesco’s writing workflow built on Spiral, Monologue, and an OpenClaw idea inbox.

“Can AI Learn Good Judgment?” by Katie Parrott/Context Window: This week, three experiments in teaching machines judgment: Dan Shipper is fine-tuning an AI copy editor on tens of thousands of editor in chief Kate Lee’s past edits to capture her calls. Head of operations Arielle Shipper turns a two-minute screen recording into reusable agent instructions, the move OpenAI just shipped as Codex’s Record & Replay. And Austin coaches Codex through tasks he can’t do himself.

🎧 🖥 “What It Will Mean to Be Human When AI Can Do Everything” by Dan Shipper/AI & I: Surge AI founder Edwin Chen—whose eval-and-data company neared $1 billion in revenue without venture funding—joins Dan to ask what motivates people once AI clears every benchmark. Chen’s view is that scaling laws suggest AI will pursue whatever goal we hand it, but the goal still has to come from a person—models have no drive of their own. 🎧 🖥 Listen on Spotify or Apple Podcasts, watch on YouTube, or follow the discussion on X.

“I Asked an AI to Audit My Own Career” by Katie Parrott/Working Overtime: Mid-quarter and unsure whether she was hitting her OKRls, Katie Parrott pointed a Codex “career coach” at her own record and had an objective read in about 10 minutes: She’d hit her goals. The piece details the setup and makes the deeper point: An agent can go hunting for evidence across Slack, Drive, and your desktop instead of relying on what you remember to mention.

Log on

Get hands-on with how Every uses AI. These are the live camps, workshops, and meetups where team members teach the workflows behind our work.

Last week’s camp

• Codex Power User Camp: A two-hour session on running Codex as a daily-driver operating system for writing, research, growth, customer support, and engineering. Watch the recording.

Upcoming events

• Q2 Demo Day (July 10): Every’s quarterly demo day, where the team shows what it shipped this quarter. RSVP.
• Every IRL (July 15): An in-person meetup at Every’s Brooklyn headquarters—check the RSVP page for time. RSVP.

From Every Studio

Compound engineering works in any coding tool

Compound engineering, Every’s approach to building software with AI agents, got a major update this week. Until now it only ran smoothly inside Claude Code; the team rebuilt it so it works the same way across other coding tools like Codex and Cursor, without the setup hassle it used to take to stay current.

It also plans projects differently now. It used to keep two separate documents—one for what you want built, one for how to build it—that fell out of sync as the build went on. Now it writes a single plan detailed enough that you can hand it to an agent and walk away. In testing, agents ran on their own for as long as six hours, building a feature, writing its tests, and opening a finished pull request with no one stepping in after the first handoff.

Cora’s rebuilt inbox moves into beta

Cora is evolving into a full-blown email client, and it’s ready for beta testers. The rebuilt version lets users read, search, compose, and reply from Cora on desktop or iPhone, alongside its existing email sorting, drafting, and twice-daily Briefs. To join the beta, email Kieran Klaassen at [email protected].

Monologue gets faster and less obtrusive

The latest updates to Monologue focus on making dictation feel nearly instant—most people will see their spoken words turn into edited text in under a second. The editing engine is also 30 percent faster without losing accuracy. Beyond speed, the updates add more transcription languages, improve how dictated text returns to other apps, and introduce a cursor indicator that automatically disappears when a user starts typing. A new Creator Mode makes Monologue easier to use during screen recordings, while new copy formats give users more options for reusing transcripts

Alignment

The worried well. I’m a Midjourney fanboy. I pay for it, I defend it at dinner against people who think AI can’t make art, and I think it makes the most beautiful images on the internet. So I am the wrong person to ask whether the AI image generator’s next idea is a good one, but given the fever pitch discussions on X last week about its latest venture, I’m going to try.

Midjourney wants to lower you into a warm pool ringed with half a million ultrasound sensors that map your whole body in 60 seconds without using radiation or the magnets commonly used in big and clunky, and sometimes scary, CT and MRI scans.

The company’s aim is to change imaging from being a dramatic event requiring a hospital visit and make it ambient and cheap, the way a blood test is. This is the kind of catch-it-early screening tool that sets the longevity crowd salivating—a path to preventing disease and, dare I say, living forever.

Doctors, however, are right to push back on the underlying premise that healthy people should be getting scanned at all. The trouble with scanning healthy people more often is that you find weird little things that may mean nothing but can’t be ignored. These are called incidentalomas, and while they’re manageable when patient volume is low, they’re not when these scans are done at a mega-scale. Midjourney’s stated ambition is to do 1 billion scans a month. If you point that many healthy, anxious people at a health system already buckling under demand, you don’t catch disease early so much as manufacture a tidal wave of follow-up scans, referrals, biopsies, and clinics clogged with people who were perfectly well until someone at a medical spa—where the scans are launching first—told them otherwise.

But it’s a charged topic, and I understand why. If a scan catches the lump that turns out to be cancer, that is worth every false alarm in the world to the person it happened to—and the testimonies you hear are always from that person. You never hear from the thousands who got a biopsy and two weeks of anxiety-inducing dread over something that turned out to be nothing.

None of this is meant to take away from what Midjourney has built. The scan gives a detailed map of fat, visceral fat, liver fat, and lean mass that can be tracked over time—very useful for GLP-1 users, who’re increasingly fixated on their fat mass to lean mass ratio.

So would I do it? For body composition, gladly. To go hunting for what might be hiding inside me, no, because I’ve seen where that road leads, and it is paved with prickly biopsies. The worried well don’t need another reason to worry; they do, however, need to know what they’re made of, and on that, the scan delivers.—Ashwin Sharma

That’s all for this week! Be sure to follow Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue. Work on documents with AI agents using Proof.

For sponsorship opportunities, reach out to [email protected].