Built on Moving Ground

On June 12, Anthropic disabled Fable 5—the most capable coding model Every had worked with—after a U.S. government ban forced the company to shut it off for everyone. There was no warning or migration window. The models we rely on, it turns out, aren’t ours to keep: A company—or a government—can switch one off overnight.

That instability is the backdrop to our week. Senior applied AI engineer Nityesh Agarwal spent weeks building around Claude’s limitations, only to watch Anthropic’s new dynamic workflows solve the problem overnight—the cost, he says, of working at the frontier. Laura Entis and Katie Parrott report on loops (where you give an agent a goal and it works in rounds, reviewing each result and re-prompting itself until the work is good enough) moving fast from engineering into non-technical work. Head of growth Austin Tedesco built an overnight NBA simulator and reworked Every’s subscription flow the same way; Katie’s half of the piece is a playbook for what to do when a frontier model disappears. Mike Taylor tried something stranger, building an AI persona of GitHub COO Kyle Daigle so the real conversation could focus on what no public record could tell him—which you can hear on this week’s AI & I. And Stella Garber, co-founder and CEO of Hoop, rebuilt an internal tool to be agent-native—then started shipping it to customers. Thursday’s edition rounds out the week with the AI topics the team has given itself permission to skip, a workflow for treating your Slack bot like a coworker, and Mistral’s surprise seat at the G-7 AI table.—Kate Lee

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

---

Knowledge base

“I Interviewed an AI Version of GitHub’s COO—Then Spoke to the Real One” by Mike Taylor/Also True for Humans: Before interviewing GitHub COO Kyle Daigle at Microsoft Build, Mike Taylor ran his questions past an AI version of Daigle built from the COO’s public record. The simulation’s misses marked where public information ran dry. He spent the real conversation on what the model couldn’t know: the three machines Daigle codes across on weekends, GitHub’s automatic model router, and an agent loop Daigle uses to grade how he communicates. Read this for the pre-interview simulation technique.

🎧 🖥 “How GitHub Deals With 17 Million Pull Requests a Month” by Mike Taylor/AI & I: With agents flooding GitHub—commits jumped from 1 billion last year to a projected 14 billion this year—guest host Mike Taylor asks COO Kyle Daigle how the platform helps developers handle the surge without telling communities which pull requests to trust. Watch or listen to this for how the home of the world’s code adapts when everyone ships with agents. 🎧 🖥 Listen on Spotify or Apple Podcasts, watch on YouTube, or follow the discussion on X.

“How Anthropic Makes Claude More Reliable” by Laura Entis/Context Window: Anthropic’s new dynamic workflows let Claude Code write its own plan and run multiple subagents through a long task. Senior applied AI engineer Nityesh Agarwal felt it directly: Weeks of custom workarounds he’d built for Claudie, Every’s AI project manager, were suddenly obsolete. Read this for the Mini-Vibe Check and to see where dynamic workflows pay off.