May 23rd, 2026

Tier 1 — Codex shipped “work while you sleep” → Always-On Reeve

OpenAI shipped the Codex overnight-autonomy stack: Goal mode (set an objective, it runs for hours/days with minimal interruptions), Appshots (Command-Command to feed any open Mac window into a thread), and locked computer use — Codex keeps driving desktop apps via a second device after your Mac is locked, controlled from your phone. Available in app, IDE extension and CLI. Verdict: verified shipped. This is the exact overnight-agent pattern Always-On Reeve Phase 2 is chasing. Action this week: install Codex, give it one real Reeve-style overnight goal (e.g. a Fillarup data refresh + morning summary), and steal the locked-use UX as the bar for Reeve’s Telegram-driven control.

Tier 1 — Anthropic’s consulting venture made its first buy → CourseBuilds

Anthropic’s still-unnamed professional-services arm made its first acquisition: Fractional AI as the operational centrepiece. Same week, Microsoft + EY announced a $1B push pairing Microsoft engineers with EY’s 400,000 consultants to get clients past “experimentation.” Verdict: verified — reported (TechCrunch/Yahoo). The labs and Big Four are both moving into bespoke AI implementation — exactly CourseBuilds’ Tier 1/Tier 2 lane. Read as validation (the demand is real and being chased by giants) and a clock: the Aria wedge is more defensible because it’s hand-built and relationship-led, but it won’t stay uncontested. Action: lock the Zaicek 30-min slot before “AI implementation partner” becomes a commodity pitch.

Tier 2 — Attention Insight → MACA’s copy/creative quality gap

Practicaly surfaced Attention Insight, an AI visual-usability checker that generates predictive attention heatmaps from a layout, colours and CTAs before any user testing — with a Figma plugin. Verdict: verified shipped (third-party). This bears directly on MACA’s two stated gaps: ad-creative quality and “does it pass human review.” Action: run one existing UBX ad creative and the campaign landing page through it — if the heatmap shows the CTA buried (a known UBX pattern), that’s a fixable, defensible data point, not a guess.

You’ve been running “Goal mode” since March

The internet spent yesterday marvelling at Codex working unattended overnight. You shipped that pattern on 2026-03-31: Always-On Reeve Phase 1 runs PaperClip as a launchd daemon (com.paperclip.server, port 3100, KeepAlive), with a 7am AEST Morning Brief cron and a 6pm AEST EOD Digest cron already firing on schedule. The brief you’re reading was written by an agent on a timer while you slept — same idea, two months earlier. The gap OpenAI closed for everyone else is persistent control after the machine locks; that’s the one piece worth lifting into Reeve Phase 2, not the concept. You’re not catching up to overnight autonomy — you’re an early adopter of your own architecture.

Printing Press — turn any API into an agent-callable CLI → MACA Performance Hub

The Rundown’s build guide walks through Printing Press, an open-source tool that generates agent-native CLI wrappers for awkward APIs (their examples: Google Flights, ESPN). npx -y @mvanhorn/printing-press install starter-pack, then the agent inspects and emits read-only commands. The angle for you: MACA’s Performance Hub already has 931 Meta API endpoints researched but no clean agent surface over them. This is a plausible 30 minutes to test whether the marketing/reviews endpoints become agent-callable without hand-building each binding. https://app.therundown.ai/guides/generate-an-agent-native-cli-from-any-api-or-website

Pieter Levels’ “audit my Mac with Claude Code”

Levels pointed Claude Code at his MacBook and it flagged FileVault off, firewall disabled, AirDrop set to Everyone, no screen lock, and SMB guest write access. You’re on a new Mac (the roy machine, all projects under ~/Developer/) — a five-minute “can you security audit my computer?” pass in Claude Code is cheap insurance before any Prevail client data lands on it. https://x.com/levelsio/status/2056705090073870460

“Google just said at I/O it’s processing 3.2 quadrillion tokens a month — 7x year-on-year — while the cost per token’s dropped about 40% since January. That’s textbook Jevons: it gets cheaper, so we use vastly more of it. Pichai’s framing was that in three years today’s AI will look as primitive as a flip phone, and engineers will manage teams of agents instead of writing code. That’s not a forecast for me — it’s the thing I’m already building Prevail around.”

Use case: Aria or Rio Tinto exec conversation where someone’s still framing AI as “a tool we’re trialling” — the quote lands the macro number, names the mechanism (Jevons, not hype), and signals you have skin in the game rather than a hot take.

Spotify just validated an audio layer for AI Edge

At its Investor Day, Spotify unveiled Personal Podcasts — AI-generated short private audio episodes tailored to your interests — plus a desktop app that builds daily audio briefings from your emails and calendar (interactivity live now; Personal Podcasts next month for US Premium). Strip the branding and that’s this brief, read aloud. AI Edge’s whole thesis is that the delivery ritual is the product — and right now the ritual is reading markdown at a desk. A TTS audio version means you consume the brief in the car or at the gym, which is exactly where you said you lose ideas (the wellness-centre origin story). Guidance: queue, don’t act. Low effort (pipe the finished brief through a TTS pass, drop an MP3 next to the Telegram ping) but it sits behind Fillarup, the UBX sale, and MACA. Park it as a one-session AI Edge Iteration-2 enhancement, not a detour this week.

• [TLDR — “Anthropic, Microsoft in talks for Maia AI chip deal”]: Covered 2026-05-22 (chip-supply talks); 30% perf claim and $5B Nov stake don’t change the read.
• [Rundown/Bagel Bots — “Anthropic to pay SpaceX/xAI ~$1.25B/month for compute”]: SpaceX $45B compute deal already covered 2026-05-22.
• [Bagel Bots — “Anthropic about to post first profitable quarter, ~$10.9B revenue”]: Covered 2026-05-22.
• [TLDR — “OpenAI Q1 revenue $5.7B, beats Anthropic”]: Covered 2026-05-22.
• [TLDR — “Cursor hits $3B ARR; SpaceX $60B option, June 12 listing”]: Incremental on $2.7B already covered; no new action.
• [Information — “Anthropic + OpenAI now 89% of AI startup revenue”]: Covered 2026-05-21.
• [TLDR — “Qwen3.7-Max agent foundation model”]: Qwen3.7 preview covered 2026-05-20; benchmark leadership noted, not yet a stack decision.
• [The Tip — “Gemini Omni Flash rolling out to all subscribers + YouTube”]: Omni announced/covered 2026-05-21; video gen isn’t on Roy’s project path.
• [TLDR — “Microsoft cancels Claude Code licenses, shifts to Copilot CLI”]: Vendor politics, fits the GitHub-eroding-lead thread already covered; no change to Roy’s Claude Code use.
• [Information — “Microsoft Opens a New Front in the Fight Over Data for AI Agents”]: Paywalled; framing-only, nothing actionable extractable.
• [Rundown — “Newsom signs first-of-its-kind AI worker-protection EO”]: US policy, not project-relevant.
• [Rundown — “Hark raises $700M at $6B for personal intelligence”]: Funding round, no near-term relevance.
• [TLDR — “Manus weighs $1B raise to unwind Meta takeover”]: China regulatory/funding drama, not actionable.
• [Practicaly — “Cursor referral: double usage for new Teams invitees”]: Promo.
• [a16z — “Peak computer science / Small Business Boom charts”]: Token-maxxing chart used above; the rest is interesting macro, not actionable this week.
• [Bagel Bots — “1-in-5 jobs high automation risk; Samsung bonuses; Nvidia concedes China to Huawei; Altman’s $2M-tokens-for-YC-equity”]: Roundup items, no project hook.
• [Bagel Bots / The Tip / Practicaly — audience-growth, client-offboarding, “Claude second brain” prompts]: Prompt-of-the-day content, not news.

• Sources scanned: 9 newsletter senders queried (Gmail); 7 fresh editions read in full
• Items extracted: ~48
• Items surfaced: 8
• Items skipped: 17
• Read time: ~5 minutes