Fable as CEO

Every’s Builder Pack now features more than $9,000 in credits and trials to our favorite AI tools. We just added two new products to the exclusive benefit for All Access members that help you and your agents differentiate your work with great design. Get two months of Paper Pro, an HTML/CSS design canvas that exports as code, and one year of Mobbin Team for up to 10 seats, with a library of more than 600,000 screens from shipped products.

Paired together, these tools give designers and generalist builders a research-to-build workflow that all connects in Builder Pack tools such as Codex, Claude Code and Cursor: Mobbin grounds agents in proven design patterns, while Paper gives them a canvas for building in code.

Hundreds of builders have joined All Access since launch, and we hosted our first office hours on Friday. Join All Access for $625 a year to claim your offers and start building something great.

Inside Every

Optimizing for collaboration

Engineer Marcus Moretti has started describing Anthropic’s models like a company org chart: “Fable is the CEO, Opus 5 is a senior engineer, Sonnet 5 is a junior engineer or analyst.”

Frontier models were once marketed as expensive generalists. Now, “we’re moving toward more of a mixture-of-models structure,” says head of platform Willie Williams. Labs are building systems of specialized models that work together inside one harness. The qualities that make a good orchestrator don’t make a good executor, and the traits that make a model good at writing don’t make it strong at debugging. So the models specialize.

Voice agents that don’t sound like a phone tree

ElevenAgents is ElevenLabs’s platform for voice and chat agents that handle service, support, and outreach. Its ExpressiveMode gives agents emotional range—warmth for a frustrated caller, urgency when the moment calls for it—so nothing sounds like a phone tree, in 70+ languages and across the tools your team already runs.

Start a free trial of ElevenLabs today

Want to sponsor Every? Click here.

“Every unit of energy spent optimizing for one thing is not a unit of energy spent optimizing for a lower-priority task,” Willie says.

Specialization also lowers costs by routing routine work to cheaper models. Just as you wouldn’t have a CEO reformat a spreadsheet, and you wouldn’t ask Fable to rename a batch of files.

Do you speak agent?

The model org chart may explain why some AI output feels as though it wasn’t written for humans. Wharton professor Ethan Mollick has watched long Fable tasks develop a distinctive dialect as agents communicate with one another, making “Claudish language ever more Claudish.” Progress reports collapse into labels, fragments, and technical shorthand.

Willie echoed Cora general manager Kieran Klaassen’s unconfirmed theory is that Anthropic trained Opus 5—an occasionally brilliant but prickly model—primarily as a subagent under Fable. If another model is the audience, pleasant prose is wasted effort. “It doesn’t matter that Opus has linguistic patterns that humans find abrasive, it mostly talks to Fable,” he says.

The problem is when that shorthand reaches a person. Head of consulting Natalia Quintero recently used Fable to analyze a large dataset. She said that its subagents’ reports read “like gibberish.”

Maybe she wasn’t the intended audience. But Fable’s draft memos were difficult to parse too.

Her fix: a handful of skills that translate agent output back into English.

“Claudish” language isn’t limited to Claude. After GPT-5.6 Sol returned an explanation he couldn’t follow, Naveen Naidu, Monologue general manager tried a lower-tech solution:

I don’t understand the issue here. Can you help me understand it?

The prompt forces the agent to explain its reasoning. Naveen’s rule: “I can outsource thinking, but I can’t outsource understanding.”

The daily driver

The models the team is using this week...

Become a paid subscriber to Every to unlock this piece and learn about:

• Which models Every team members are using this week
• How Arielle and Austin’s Codex agents collaborate with each other on site metrics
• A prompt to create your own company context packet for colleagues