06-reference

every built on moving ground

2026-06-21·reference·source: Every·by Kate Lee (editorial intro); Laura Entis, Mike Taylor, Stella Garber, Katie Parrott, Ashwin Sharma (contributors)
ai-agentsfrontier-modelsagent-native-toolingplatform-risksoftware-development

"Built on Moving Ground" — Every (Context Window digest, June 21 2026)

Why this is in the vault

This issue lands directly on RDCO's operational reality: what happens when the frontier model you depend on disappears overnight, and how teams are building agent-native workflows without getting caught flat-footed by platform volatility.

Issue contents

Editorial frame (Kate Lee): The week's theme is triggered by Anthropic disabling Fable 5 on June 12 with no migration window due to a U.S. government ban. Every frames the full issue around the instability of building on models you don't control.

Articles covered:

  1. "I Interviewed an AI Version of GitHub's COO—Then Spoke to the Real One" (Mike Taylor / Also True for Humans) — Taylor built an AI persona of GitHub COO Kyle Daigle from public record, ran interview questions through it first, then used the simulation's blind spots to focus the real conversation. GitHub commits projected to jump from 1B to 14B this year as agents flood the platform. Daigle uses an agent loop to grade his own communication.

  2. "How Anthropic Makes Claude More Reliable" (Laura Entis / Context Window) — Covers Anthropic's dynamic workflows: Claude Code writes its own plan and runs subagents through long tasks. Every's own Claudie (AI project manager) had workarounds built by hand — Anthropic's feature made them instantly obsolete. Includes "Steal This Workflow" (treating your Slack bot like a coworker) and mini vibe checks.

  3. "Loops for Non-Coders" (Laura Entis + Katie Parrott / Context Window) — The agent loop pattern (hit stopping point → write new prompt → continue) has crossed into non-technical work. Every's head of growth used loops to build an overnight NBA simulator and rework the subscription flow. Parrott adds a frontier-model risk playbook: save sessions, build while the window is open, audit which tasks actually needed the frontier model vs. could run on a cheaper one.

  4. "We Built Our Own Agent-native Tool" (Stella Garber / Every) — Hoop's non-engineer co-founders built an internal AI tool to organize scattered customer-call notes. Architecture: give the model tools and let it reason inside Slack rather than scripting fixed prompt sequences. Now shipping to customers.

  5. "How GitHub Deals With 17 Million Pull Requests a Month" (Mike Taylor / AI & I, podcast) — Companion audio to the COO interview.

ALIGNMENT (Ashwin Sharma): Short personal essay prompted by Juneteenth and James Baldwin's Go Tell It on the Mountain. A groom's AI-written wedding speech is the cautionary counterpoint; the thesis is that specificity and truth are what surprise — and AI can't supply those for you.

From Every Studio:

Mapping against Ray Data Co

Platform risk / model continuity — The Fable 5 overnight kill is a direct stress test for any founder running a COO-agent harness. RDCO's own session-management SOP and the subagent dispatch pattern exist partly to stay resilient to exactly this. The Parrott playbook (save sessions, audit which tasks need frontier vs. capable-enough models) is worth reviewing against RDCO's skill stack.

Agent-native build pattern — Stella Garber's Hoop walkthrough (tools + reasoning > fixed prompt scripts) maps exactly onto how RDCO's skills are structured. Validates the thin-harness / fat-skills architecture decision.

Non-engineer loops — The Austin Tedesco NBA simulator and pricing-page loop story is a signal that "loop" tooling is going mainstream-ish. RDCO's /loop skill is ahead of this curve; worth watching as a potential content-as-product angle for Sanity Check.

Claude Code subagents — The Nityesh Agarwal / Claudie story (hand-rolled workarounds made obsolete by Anthropic's feature) is a cautionary tale about over-engineering at the frontier layer. Relevant to how RDCO builds skills vs. waiting for native capabilities.

phData DSA context — GitHub's 14B projected commits / agent-flood framing is good data for client conversations about why AI-native dev practices matter now, not in two years.

Related