June 18, 2026
Cole Medin (creator of the Archon open-source project) joins Nate Herk for an in-depth session covering the full stack of directing Claude Code as a business operating system — from the Plan-Build-Verify-Evolve loop to adversarial development, the RALF pattern, and hooks-driven session memory.[1]Nate Herk | AI Automation — How to Build Effective Claude Code Agents in 2026 The central thesis: stop vibe coding and start directing — Claude Code is most powerful when you treat it like a co-founder that needs intent, not just instructions.
Cole distinguishes "vibe coding" (prompt and pray) from directing: a director gives Claude the why behind every task, provides structured context, and builds systems that make future work more reliable. Claude Code should be treated as a co-founder or AI operating system for the entire business, not just a coding helper.
A four-step framework for repeatable results: (1) Plan — use sub-agents to research options before writing a spec, ensuring both human and agent alignment. (2) Build — implement with the spec as context. (3) Verify — automated validation checks (rendering Excalidraw diagrams, browser automation, etc.). (4) Evolve — every bug becomes a permanent upgrade to CLAUDE.md or skills. Cole prefers a custom planning skill over Claude's built-in plan mode for full structural control.
The "agent harness" is the wrapper of tools, context, and system prompts around the base LLM. Crucially, for Opus 4, performance degrades significantly past ~250K tokens — the "dumb zone." Cole chains multiple Claude Code sessions together (planner → implementer → verifier) to avoid this threshold and keep context windows clean.
RALF is a multi-agent orchestration pattern where a main session defines task phases, then individual agent sessions handle each phase sequentially. Unlike Claude's built-in agent teams feature (which Cole finds token-heavy and unrefined for deep development), RALF uses code-driven orchestration to maintain determinism.
Cole's open-source Archon project makes multi-agent workflows more deterministic by using code (not another LLM) to decide when agents run and in what sequence. This avoids the compounding unpredictability of LLM-to-LLM orchestration for production systems.
Claude Code hooks run code before any tool invocation. Cole uses them for two purposes: (1) Security — blocking forbidden file paths, dangerous commands, or database writes. Mental model: assume the agent will do anything it can touch, so scope API keys and MCP permissions narrowly. (2) Session memory — automatically summarizing completed sessions into daily logs, then running a nightly "dreaming" process that promotes key decisions to long-term memory.
A separate Claude Code session is prompted to aggressively critique and find flaws in the primary session's output. This sidesteps LLM sycophancy — instead of asking for opinions, you run empirical adversarial tests. A code review skill that asks "what could go wrong?" and actively invokes edge cases complements this.
Providing the rationale behind a request — not just the spec — significantly improves output quality by shaping the model's approach through intent. Cole's top-3 Claude Code features: (1) Skills (most powerful, fully customizable reusable prompts), (2) Hooks (security + session memory automation), (3) Sub-agents for parallel research. The skill+CLI combo is the most token-efficient alternative to MCP servers.
Theo documents his conversion to agent loops: instead of manually prompting each step, you design automated systems where agents prompt themselves, write throwaway code to trigger sub-agents, and spawn dynamic workflows tailored to each problem.[2]Theo — I guess we're writing loops now? His most striking demo: a single message created a heartbeat-based workflow that implemented, reviewed, iterated, and merged four stacked PRs while he slept.[3]Theo — Prompt Loops, Not Individual Instructions
Theo was initially skeptical of agent loops but changed his mind after Codex's thread-spawning-thread capability demonstrated that models writing code to orchestrate their own next steps is qualitatively different from traditional programming. The unlock: 240 lines of throwaway code executed exactly once to trigger sub-prompting — code as an intermediate step between model runs.
Hardcoding agent personas (adversarial reviewer, security reviewer, etc.) misses the point — agents should build their own context dynamically based on the actual problem rather than being assigned static roles. Loops that generate workflow structure tailored to each problem beat rigid agile sprints.
"Loops created loops" — the meta-loop spawned sub-loops dynamically, with the agent adapting its own orchestration strategy mid-run.
Theo's centerpiece experiment: a single message at 2:29am created a dynamic heartbeat-based workflow. By 6:50am the agent had implemented, reviewed, iterated, and merged four stacked PRs — all while Theo slept. The meta-loop spawned sub-loops dynamically; the agent adapted its orchestration strategy mid-run.
Map every manual step you take after an agent finishes — running servers, testing, committing, pushing, filing PRs — and delegate those steps back to the loop. The /goal primitive enables single-thread infinite loops. Cost warning: 3M tokens for just three code review comments. However, Theo's actual June usage was ~$10,000 of inference on $600 in subscriptions (three $200/month plans), suggesting subscription users have enormous untapped headroom.
AI Jason introduces "loop engineering" as a distinct discipline from prompt engineering — designing automated systems that trigger coding agents rather than manually prompting them.[4]AI Jason — wtf is Loop Engineer & how to setup for real The key architectural insight: multiple agent loops sharing a common "signals" file system compound each other's value — a support loop, SEO loop, and ads loop all writing to shared signals can learn from and reinforce each other.
Loop engineering is the next evolution after prompt engineering: instead of crafting individual prompts, you design the triggers (cron jobs, other agents, webhooks), orchestration logic, and environment that make agents run autonomously. The "agent harness" is everything non-model — prompts, hooks, tools, and context.
A codebase must be legible (CLAUDE.md + custom lints so agents know what to change), executable (a dev.local script + worktree support so agents can spin up without token overhead), and verifiable (Playwright CLI + a separate read-only verifier agent that checks work independently).
Loop engineers maintain three layers: (1) Artifacts — typed output folders (docs, signals, tasks) with schemas shared across loops. (2) Loop contracts — per-loop README files with goal, workflow, backlog, and timeline. (3) Global work-log.md — a cross-loop journal of what got done and why. The signals folder is where loops leave intelligence for each other.
Setup steps: create skills for tool access (Intercom, Stripe, Supabase), generate a business context file in CLAUDE.md, scaffold the artifacts structure, run a manual test, then commit to an automated loop contract. This support loop then runs on triggers, writes to the shared signals folder, and feeds SEO and ads loops with customer language.
AI infrastructure spending now comprises 39% of marginal U.S. GDP growth — larger than tech's share at the dot-com peak. Ex-AI, H1 2025 growth would have been 0.1%.[5]The AI Daily Brief — Why Only AI Training Can Save the Economy The episode argues this structural dependence creates a single lever to sustain the AI economy: mass-scale AI training — which currently suffers an "abysmal market failure" with only 28% of organizations having changed business processes despite widespread AI awareness.
AI investment = 39% of marginal GDP growth (vs. 28% for tech at dot-com peak). $800B in big-tech capex in 2026 — David Sacks estimated it at a 2.5–3% GDP tailwind. Ex-AI, first-half 2025 growth would have been 0.1%. This is no longer a sector story; it's the entire growth story.
2024 "AI bubble" fears were driven by seat-based math ($20–$200/month per user couldn't justify trillions in capex). The shift to agentic, usage-based consumption changes the equation: Anthropic's run rate grew from $30B → $47B as Claude Code drove per-person AI economics to thousands per month. Enterprises with $1M+ annual spend went from 500 to 1,000+ in two months.
Enterprise spending caps shape what gets attempted, creating a "known ROI bias" — companies deploy AI for incremental productivity gains and away from the experiments that would unlock transformational value. Real examples: Uber's $1,500/month cap, Factory's model routing saving $13M over 30 days. Every AI business is now a token efficiency business: DeepSeek/Kimi adoption, Cursor Composer 2.5 at "1/10th cost."
EY data: only 28% of organizations changed business processes despite widespread AI awareness. DataCamp: "awareness without confidence." Managing agents is a new knowledge-work primitive closer to management training than software engineering. The episode's call to action: AI labs must lead mass-scale enablement — their own survival depends on it, because token growth is the public-market pressure they face post-IPO, and CFO scrutiny will cap that growth unless workers actually know how to generate value with AI agents.
Anjney Midha (AMP founder, ex-a16z) runs an independent compute grid (~1.3 GW base load, targeting ~6 GW spike) plus a capital arm ("Foundry") that backs Anthropic, Mistral, and others — giving him a unique vantage on why well-resourced AI labs underperform.[6]Latent Space — Why AI Labs With Unlimited GPUs Still Fail His central thesis: compute and capital are no longer the binding constraints — culture, mission alignment, and disciplined "output maxing" are.
Node utilization below 95% is effectively an outage. Model flop utilization (MFU) typically runs at only 60–70%. Most labs don't publish MFU because it's embarrassing. The failure mode: scaling faster than you can run iterative bring-ups, combined with frayed internal culture.
"From an engineering perspective it's basically output maxing. It's the Department of Output Maxing."
Small misalignments compound across the capital→cluster→output chain: a one-degree error at capital allocation becomes catastrophic drift by the time it reaches model outputs. This explains why abundance — plentiful compute and money — can actually increase failure rates by removing the scarcity-driven discipline that hardship forges.
Anjney attributes Google's relative miss partly to its internal credit-bidding compute marketplace — a market mechanism that allocates GPU time to the highest internal bidder. This creates adverse selection: the teams that "win" compute aren't necessarily the ones with the best P0, just the ones with the biggest political capital. DeepMind's research embargo (publishing only what's not good enough to keep internal) creates a similar adverse selection problem for the broader ecosystem.
"The stuff that gets published is the stuff that's not good enough. So there's an adverse selection problem."
Anthropic's coding → AGI thesis is a sharp, defensible P0. Researchers as CEOs (Anastasios Antoniadis, Dario Amodei) means the people closest to the science are making allocation decisions. Culture is "a set of actions, not beliefs" — if you stop doing the daily actions that demonstrate it, the culture disappears within months.
"Culture is not a set of beliefs. It's a set of actions. If you stop taking the actions that demonstrate the culture, it disappears."
"Anthropic has been the most prepared company for four years. Luck favors a prepared mind."
AMP operates multi-cloud, multi-silicon compute with net-positive data centers as a community engagement strategy. "Neoclouds" is marketing speak — there are really just well-run and poorly-run versions of the same business. Foundry co-invests alongside AMP's compute in frontier labs (Anthropic, Mistral, Black Forest Labs, Periodic), creating aligned incentives across the capital→cluster→output chain.
Sandipan Bhaumik (Databricks technical lead) presents a five-pillar framework — evaluation, observability, data foundation, multi-agent orchestration, and governance — as the prerequisites before writing production agent code.[7]AI Engineer — The Production AI Playbook: Deploying Agents at Enterprise Scale The case study: a retail bank rebuilt an $85K failed POC with an eval-first approach (model chosen in week 7 of 8), catching a stale-embedding RAG incident post-launch via CSAT drops rather than by chance.
Most enterprise AI projects fail not from bad models but from three missing capabilities: observability (no visibility into what the agent is doing), evaluation (no system for measuring quality), and governance (no control over who can do what with which model). Teams choose the model first, then discover these gaps at production scale.
Deterministic (regex, classic ML) — fastest, most reliable. Semantic (LLM-as-judge) — catches meaning-level errors. Behavioral (tool call correctness, loop detection) — the most commonly skipped layer, yet the most critical for agents. Behavioral evals against large datasets are expensive; run a subset in CI and the full suite only on merge to main. Treat eval datasets as living systems with structured commit messages for prompt changes.
Distinguish "question data" (RAG inputs, API data) from "tracking data" (traces, which need their own schema strategy). Getting the data foundation right takes 60% of project time in practice but is almost never planned for.
Three orchestration patterns: (1) orchestrator-worker, (2) choreography via message bus, (3) human-in-the-loop. Production incident playbook: detect (eval dashboard / CSAT drop) → diagnose (trace inspection) → contain (prompt rollback or circuit breaker) → fix (update eval suite so it catches the failure class going forward).
ZAI launched Z-Code alongside GLM-5.2 — a free coding agent (analogous to OpenAI's Codex) with 5M daily free tokens, fine-tuned specifically for GLM.[8]AICodeKing — FULLY FREE GLM-5.2 + Z-Code: This is ACTUALLY GOOD! The headline benchmark: GLM-5.2 scores 74 on Frontier SWE (vs Opus 4.8 at 75, GPT-5.5 at 72), more than doubling its predecessor GLM-5.1 which scored just 30 — while being MIT-licensed and open-weights.
Z-Code is ZAI's free coding agent interface for GLM models. It offers 5M free daily tokens, a clean task-oriented interface, and tight GLM fine-tuning. Current limitations: no file explorer panel, no Git init support, and the browser panel isn't agent-controllable (agents can view but not interact with it). Despite these gaps, the free tier dramatically changes the access equation for GLM models.
GLM-5.2 on the Artificial Analysis Intelligence Index: 51. MIT-licensed. Terminal Bench: 81 (vs Opus 85). SWE-Bench Pro: 62.1. PostTrain Bench: 34.3. MCP Atlas tool use: 76.8 (near-parity with Opus at 77). Tool Decathlon: 48 (trails Opus at 59, GPT at 55). Strong general knowledge: AN2026 99.2. All long-horizon evals used 1M token context with max effort.
Frontier SWE: 74 vs Opus 4.8 at 75 and GPT-5.5 at 72. The prior GLM-5.1 scored just 30 on this benchmark — GLM-5.2 more than doubled it. SWE-Bench Ultra: 13 (trails Opus at 26, but a meaningful open-weights presence in a benchmark the model family was previously absent from). The combination of near-frontier long-horizon coding ability with MIT licensing and a free usage tier is the differentiator.
Artificial Analysis launched AA-Briefcase, a benchmark evaluating AI on realistic multi-week knowledge work projects across Data Science, Product Management, Banking Operations, and Heavy Industry Strategy — 91 tasks with thousands of input files per task.[9]Artificial Analysis — Announcing AA-Briefcase Claude Fable 5 leads overall, yet even it achieves 100% rubric pass rates on only 3% of tasks; on 31 of 91 tasks no model scores above 50%.
Tasks simulate actual multi-week professional projects, requiring models to process hundreds of input files (Slack threads, emails, company documents, meeting transcripts). Graded on three dimensions: rubric checks (objective correctness), analytical quality (rigor via pairwise comparison), and presentation quality (professional polish). This is deliberately harder than SWE-bench — it models knowledge work, not just code.
Cost varies by over 800x per task: from $31+ (Claude Fable 5) to $0.04 (DeepSeek V4 Flash). A strong behavioral signal emerged: leading models visually inspect their outputs 12–21 times per task versus ~0–2 times for weaker performers. Lower-capability models fail at execution; advanced models struggle with hidden requirements and incomplete analysis — a qualitatively different failure mode.
Anthropic's Project Fetch Phase 2 found Claude Opus 4.7 operating fully autonomously completed robotic quadruped tasks over 37x faster than human teams on four comparable tasks — while writing approximately 10x less code.[10]Anthropic Research — Project Fetch: Phase two The remaining hard limit: fine-grained ball retrieval requiring continuous visual feedback, which experienced human roboticists could accomplish but Opus 4.7 could not.
Opus 4.7 ran with adaptive thinking at maximum effort inside Claude Code, with human involvement limited to laptop connection and command approval. Tasks included connecting to sensors, writing control programs, detecting objects, and autonomous retrieval. Three experimental trials measured completion time and success rates against human teams (one with Claude access, one without).
The dramatic code efficiency — 10x less code than the Claude-assisted human team — suggests the model found more direct solution paths rather than accumulating scaffolding. The visual feedback limitation (continuous sensorimotor adjustment) represents the frontier of what current LLMs can do with robot control. The research frames AI as transitioning toward reliable physical task automation for discrete, well-defined subtasks.
The Batch Issue 358 reports that independent evaluators of Claude Fable 5 encountered a significant transparency problem: safety classifiers automatically downgrade or refuse flagged prompts without clearly notifying evaluators, making consistent measurement impossible.[11]The Batch Issue 358 — Testing Mythos and Fable, Moving Beyond SWE-bench The model refused ~8% of economically useful benchmark tasks; on science questions, accuracy dropped from 93% to 56% when refusals were counted as failures.
The core issue: evaluators can't reliably tell when a lower score reflects model capability vs. a safety classifier intervention. This creates inconsistent measurement across testing interfaces and populations. Andrew Ng's letter in the same issue frames the broader context: Anthropic's U.S. government-directed restrictions on Fable 5 demonstrate how corporate and governmental power over model access is accelerating global pushes toward sovereign AI and open-source alternatives.
The Batch highlights three new software engineering benchmarks designed to move evaluation beyond SWE-bench's single bug-fix frame.[11]The Batch Issue 358 Results: DeepSWE has GPT-5.5 at 70% and Claude Opus 4.8 at 58%; ProgramBench found no model could pass all tests (Opus 4.7 reproduced only 3% of programs); ITBench-AA put Opus 4.7 at 46.7% full recall on infrastructure diagnosis.
The Carnegie Mellon POPE paper (in the same Batch issue) showed that providing solution prefixes during RL training significantly improves performance: AIME 2025 improved from 49.6% to 53.1% pass@1; HMMT 2025 from 31.0% to 37.8%.
Nvidia released Nemotron 3 Ultra — a 550-billion-parameter hybrid transformer-mamba architecture model optimized for long-context agentic tasks.[11]The Batch Issue 358 It's the fastest among comparable open-weights models at ~183 tokens/second, scores 47.7–48.2 on the Artificial Analysis Intelligence Index, and is fully open: weights, training data, and code all freely available.
The hybrid transformer-mamba architecture offers better throughput than pure transformer models at long context lengths — relevant for agentic workflows that accumulate large context windows. Fully open weights (including training data and code) make it the most accessible frontier-adjacent model in the open-weights ecosystem. At 47.7–48.2 on the AA Intelligence Index, it sits well below Fable 5 but near the top of the open-weights tier.
Agent-Reach is a 28K-star MIT-licensed Python CLI/library that installs in one natural-language sentence and registers itself as a capability layer for coding agents — enabling multi-platform internet research across Twitter/X, Reddit, GitHub, Bilibili, Xiaohongshu, and more.[12]Better Stack — Your AI Agent Is Missing Half the Internet… Until Now (Agent-Reach) The key differentiator: it covers Chinese developer communities that English-centric tools ignore.
Uses a "platform channels" architecture with primary and fallback backends per platform — so if one data source is unavailable, it automatically degrades to an alternative. Each platform has its own extraction strategy rather than a generic scraper.
A Claude Code agent installed Agent-Reach via a single natural-language prompt in VS Code, then successfully performed multi-platform research across Twitter/X and Bilibili, returning structured results with source links. The Chinese platform coverage was the most distinctive result — it found conversations and perspectives that English-only tools would have missed entirely.
Best for read/search/extract use cases. Not a browser automation tool — can't handle complex multi-step UI interactions. For standard web pages without login requirements, Firecrawl is simpler. Agent-Reach's value is specifically for social platforms and structured community data.
OpenAI released Record & Replay for Codex — you demonstrate a multi-step workflow once (e.g., uploading a YouTube video: pulling metadata from a spreadsheet, adding thumbnail, captions, saving as private), and Codex records it and converts the recording into a reusable skill.[13]OpenAI — Record & Replay in Codex Future runs of the same workflow require only attaching the relevant files — Codex executes autonomously via computer use, browser use, or connected plugins.
The skill captures where metadata lives, how the upload package is organized, and how to verify each step. This extends Codex from a code-generation tool to a general-purpose workflow automation system. Use cases demonstrated: video publishing, pull request formatting, calendar invite setup — essentially any repeatable workflow a human currently does step-by-step. The model doesn't need a step-by-step prompt on replay; it references the learned skill.
OpenAI released a brand documentary showing hundreds of practicing physicians — pediatricians, psychiatrists, hematologists, and specialists — evaluating ChatGPT health responses for accuracy and impact in real time.[14]OpenAI — Improving health intelligence in ChatGPT The team's stated mission: democratize super-specialty medical knowledge globally by making it accessible where specialists can't physically be.
Physicians evaluate both accuracy and the likely real-world impact of a response on the person receiving it. A hematologist and transfusion medicine specialist described the vision: take expert knowledge "to places where I can't go, but maybe an AI model can." The production-facing doctors see 15 patients per day themselves; through ChatGPT health, they aim to influence millions. The documentary highlights the "how you say it matters as much as what you say" principle guiding response calibration.
Prefect's Horizon platform adds role-based access control and governance to MCP servers — solving the enterprise problem of 100+ MCP servers with no visibility into who has access to what, which creates SOC 2 compliance gaps.[15]Prefect — Running MCP with no gateway? Crazy. About 70% of Horizon's enterprise customers arrive with unmanaged MCP environments; the platform consolidates them into one governed gateway.
Raw MCP means everyone has access to everything by default. Horizon adds an auth layer and RBAC so organizations can say "this tool is available to the marketing team but not engineering" — without writing custom auth. SOC 2 type 2 certification requires knowing where data went; unmanaged MCP makes that impossible. The practical use case shown: three separate Gong MCP instances (per team) consolidated into one with per-user permissions, eliminating redundant API key management and enabling centralized audit logging.
Simon Willison shipped Datasette Apps — self-contained HTML+JavaScript applications running in sandboxed iframes within Datasette instances, able to execute read-only SQL queries via JavaScript and write via pre-configured stored queries.[16]Simon Willison — Datasette Apps The key distinction from Claude Artifacts: these apps have access to persistent relational databases. Companion release datasette-acl 0.6a0 adds fine-grained multi-user permission management.[17]Simon Willison — datasette-acl 0.6a0
Security is layered: iframes with sandbox="allow-scripts allow-forms" prevent cookie/localStorage access; an injected Content Security Policy blocks external network requests and data exfiltration; MessageChannel() handles locked-down communication between the app and parent frame. CSP domain allowlisting is restricted to trusted administrators. The system is AI-friendly: copyable prompts for LLM-assisted app development, and integration with the Datasette Agent for Claude-style artifact creation inside a Datasette instance.
Two new developer tools launched on June 18: Eve (Vercel) — an open-source agent framework that organizes your entire AI assistant stack as a directory of files (markdown for instructions, TypeScript for tools), with sandboxed compute and one-command Vercel deployment[18]Github Awesome — Eve: Vercel-backed agent framework; and Junction — a VS Code sidebar that embeds local AI coding agents directly next to your code, eliminating context switching between terminal and editor.[19]Github Awesome — Junction: VS Code sidebar for coding agents
Eve maps the entire agent stack to a folder: markdown files for instructions, TypeScript files for custom tools. Sessions survive crashes and redeploys via durable state. Local development deploys to Vercel functions in one shot. The file-based model makes agent configuration version-controllable and inspectable.
Junction adds a chat sidebar to VS Code for local coding agents — you point it at files, changes land in your project in real time, and the complete agent loop (prompt, file changes, verification) runs inside the IDE without switching windows.
The Mythos/Fable standoff escalated: Commerce Secretary Lutnick explicitly warned against distributing these models to foreign persons, the restriction list expanded to include companies with potential China ties, and internal Anthropic employees expressed frustration about the lack of transparency.[20]The Rundown AI — Inside the deadlock keeping Mythos offline In parallel, G7 leaders met with Dario Amodei and Demis Hassabis in France, where they proposed a "U.S.-led AI coalition" addressing model access, chip exports, and safety.[11]The Batch — Issue 358
Andrew Ng's Batch editorial frames the broader consequence: the U.S. government's use of Commerce Department authority to restrict Fable exports has accelerated global pushes toward sovereign AI development and open-source alternatives, as nations worldwide seek to reduce dependency on U.S.-controlled model access.
Supporting data points from the same Rundown issue: Pew Research found ~50% of U.S. adults now use AI chatbots, but nearly 40% expect AI to harm society over 20 years — with ChatGPT at 44% recognition vs Claude's 6%. An Anthropic study of 400K Claude Code sessions found domain expertise matters more than coding skill; users made 70% of planning decisions while Claude handled 80% of execution. OpenAI Q1 2026: $5.7B revenue and $3.7B cash burn — both tripling year-over-year.
Tim Cook confirmed Apple price increases are "unavoidable" as AI demand has driven memory chip prices up 90% since October 2025 — DRAM costs per iPhone device are projected to rise from ~$39 to ~$145, pushing the iPhone 18 Pro starting price from $1,099 to $1,299 with manufacturing costs up from $582 to $726.[21]Tech Brew — Hey Siri, how much is an iPhone now?
The structural irony: AI features (a revamped Siri, on-device inference) are both the reason iPhones need more memory and the reason Apple hopes customers will pay more. Apple maintained a 47% profit margin on iPhone 17 Pro and held ~7 in 10 U.S. smartphone sales in Q4 2025, giving it pricing power to pass costs on. Price increases are expected across MacBooks, iPads, and other products when the iPhone 18 Pro, Pro Max, and Apple's first foldable phone arrive this fall. Apple may eventually charge separately for enhanced Siri functionality.
Morning Brew's June 18 edition covers two macro developments: a tentative U.S.-Iran agreement was released following the Strait of Hormuz ceasefire deal reached June 15, and the Federal Reserve signaled rates may rise at Kevin Warsh's debut FOMC meeting as new Fed Chair — a hawkish signal marking a potential policy shift.[22]Morning Brew — June 18, 2026
The tentative US-Iran agreement follows the ceasefire deal that reopened the Strait of Hormuz on June 15 — a geopolitically significant waterway for global oil supply chains. Warsh's debut Fed meeting as chair signals a shift from the rate-hold posture of recent months. Carvana also announced plans to introduce new cars on its traditionally used-car platform — a strategic expansion into new vehicle sales.
Google DeepMind's Logan Kilpatrick argues that as AI-generated content floods every platform, authenticity — your own voice, words, and presence — becomes a competitive advantage, not a limitation.[23]Sequoia Capital — Logan Kilpatrick on using AI He uses Gemini Omni enthusiastically for exactly this reason: it changes the production context around authentic content (the set, the format, the frame) without replacing the person.
"There's just so much alpha in authenticity."
Kilpatrick doesn't use AI to write any content he produces personally. His vision for generative media: "amplified authenticity" rather than AI avatars — the original content, the original person, just in a different and more interesting context. Gemini Omni is his ideal example of this distinction: "It's the original content. It's the person. It's like the personhood is there. It's just different and amplified."
Gergely Orosz argues that for any production system with database state, rolling back code is dangerously misleading: schema changes mean reverting to v1 after a v2 bug could create a version mismatch that's worse than the original bug.[24]The Pragmatic Engineer — Rollbacks are a lie The correct model is always "roll forward" — ship v3 with the fix as fast as possible.
Most teams operate on "sheer luck" that they've never hit a rollback + schema mismatch catastrophe. The advice: stop framing deployments as reversible; instead build for fast-forward recovery. If v2 is broken, the answer is a v3 with the fix in it, deployed as quickly as possible — not restoring v1 code against a v2 schema. This reframing also changes how teams design migrations and data changes.
Historian Sarah Paine appeared on the Dwarkesh Podcast to dissect Putin's geopolitical operating manual: no two-front wars, no great-power neighbors, take neighbors sequentially through destabilization and absorption, use buffer zones, and play the "jackal state" — moving in to claim kills that weakened opponents made.[25]Dwarkesh Patel — Putin's playbook - Sarah Paine
Paine notes the structural flaws in this strategy: you end up surrounded by dysfunctional states (which Russia and China are, and which may not be entirely accidental), you can't build enduring alliances (neighbors eventually identify the hegemonic threat), and there's no built-in council for when to stop expanding. Historically, both Russia and China have repeatedly overextended — accounting for the implosions of various empires and dynasties across their long histories.