May 26th, 2026

Tier 1 — MCP next-spec release candidate drops, with breaking changes [Ben, MetaAdsMCP, Power Platform docs tool]

The 2026-07-28 Model Context Protocol spec is now a public release candidate — the biggest revision since launch. It introduces a stateless core that scales on ordinary HTTP, OAuth/OpenID-aligned auth, a formal deprecation policy, and explicit breaking changes. Verdict: verified shipped (RC live; final ships July 28). This hits three of your builds directly: Ben (Xero + Google Workspace MCP — you already migrated Workspace from stdio to shared HTTP, so the stateless direction is tailwind), MetaAdsMCP (Cloudflare Workers, stateless-friendly), and the Power Platform docs tool. Action this week: read the breaking-changes list and audit Ben’s MCP clients now, not in July.

Tier 1 — Google rebuilds Search/AI Mode ads with Gemini [MACA, Prevail paid search]

Google launched Gemini-built ad formats: Conversational Discovery Ads (creative generated to answer the exact query), Highlighted Answers (ads inside AI Mode recommendation lists), live chat agents in-ad, and AI-Powered Shopping Ads with model-written explainers — all “Sponsored”-labelled. Verdict: research preview / rolling out (live testing now, full rollout “coming months”). Where Google goes on AI-native ad units, Meta follows — this is the shape MACA’s output has to survive. Action: study the Conversational Discovery mechanic; MACA’s copy now has to read as a tailored answer to intent, not a static headline.

Tier 1 — DeepSeek makes its 75% V4-Pro cut permanent; Gemini 3.5 Flash Low lands [MACA copy cost, Ben, Fillarup]

DeepSeek permanently cut V4-Pro to $0.435/M input, $0.87/M output — tops Artificial Analysis “intelligence-per-dollar,” reportedly 12–19× cheaper than frontier rivals at comparable quality. Same day, Gemini 3.5 Flash (Low) shipped: ~45% fewer tokens, beats Flash (High) on SWE tasks. Verdict: verified shipped. Your cost-exposed loops — MACA’s multi-agent copy generation, Ben’s daily scans — are exactly where intelligence-per-dollar matters. Action: run a cost/quality bake-off of MACA copy gen against V4-Pro and Flash Low before the next pipeline run; log it in api/lib/costs.ts.

The Rundown’s “build an AI secretary that plans your day” is Reeve’s Morning Brief, shipped two months ago

The Rundown ran a step-by-step guide today teaching readers to use Codex/Claude Code to scan Slack, Gmail and calendar each morning and write a prioritised MonoNote.md to-do list “that gets better every day.” That’s Always-On Reeve’s Morning Brief routine — running on a com.paperclip.server launchd daemon (port 3100, KeepAlive) on a 7am AEST cron since 2026-03-31, with the EOD Digest as the evening counterpart. The newsletter is pitching as a weekend project the exact pattern you’ve had in production for eight weeks. You’re not behind; you’re the case study.

Anthropic’s coming “Memory Files” is the structure you already run

TLDR flags an upcoming Claude Memory update — “Memory Files” that distribute notes across multiple structured documents organised by topic, project, and context. That is precisely your operating model: active-projects.md, covered-stories.md, ~/Reeve/learnings/, the memory/ files. This very brief uses covered-stories.md as structured, de-duplicating memory. Anthropic is productising the file-per-context pattern you’ve been running by hand — validation that the architecture was right.

OpenAI — Evaluating multi-agent systems at scale

A macro-evaluation workflow for agentic systems that analyses patterns across entire populations of traces rather than isolated failures. (TLDR link — search “Evaluating multi-agent systems at scale”, OpenAI.) Your angle: MACA is a 14-agent / 4-wave pipeline and your honest gap is copy quality you can’t yet reliably predict. Trace-population evals are how you stop debugging one bad ad at a time and start finding systemic copy-quality failure modes across runs. 30 minutes here could be the difference between “feels better this session” and a measurable quality gate.

Practicaly — connect Microsoft Clarity to Claude for behaviour-driven conversion fixes

Claude now reads Microsoft Clarity (free) as a connector and returns concrete UX/conversion changes — one example chewed through 100k+ sessions and 20k signups and came back with specific recommendations. (Practicaly). Your angle: Fillarup and CartQuote are both pre/early-launch where conversion is everything and you have no analyst. Wire Clarity into Fillarup at launch and you get an on-demand conversion analyst for $0 — directly relevant to the “collect first 20 reviews / iterate” CartQuote next-step.

“McKinsey is rethinking the billable hour right now — the FT reported it this week — because AI is collapsing the value of the hours, and clients want fees tied to outcomes instead. That’s not a McKinsey problem, it’s the whole consulting model repricing. It’s exactly why I’m not selling AI training by the seat or the hour — the value is the embedded outcome, the workflow your team actually uses on Monday, not the days I spend in the room.”

Use case: Drop this with Michael Zaicek when CourseBuilds activates, or in any Aria/RT/AI-pro conversation about what AI does to professional-services pricing. It signals you see where consulting economics are heading (McKinsey, EY+Microsoft, the Anthropic/Blackstone venture all moving the same way) and that CourseBuilds’ outcome-based pricing is a deliberate read of that shift, not a number you picked.

(Nothing clears the first-mover bar today.)

• [TLDR — “Anthropic prepares Mythos 1 for Claude Code and Claude Security”]: Mythos covered repeatedly; no new action for your stack until general release.
• [Rundown / TLDR — “Claude Mythos finds 10,000+ critical vulnerabilities (Project Glasswing)”]: Impressive (Cloudflare 2k bugs, Mozilla 271, a bank blocked a $1.5M wire), but defensive-security niche, not your wheelhouse.
• [TLDR — “The Neocloud Boom”]: SpaceX renting Colossus to Anthropic at $15B/yr, $7.5T buildout — macro colour only; SpaceX–Anthropic compute already covered.
• [TLDR — “Anthropic’s march to profitability”]: $10.9B Q2, compute 71c→56c per revenue dollar, Oct IPO — already surfaced May 22 and 25.
• [Rundown / Practicaly — “Perplexity open-sources Bumblebee” / “Claude security-audit your computer”]: Pieter Levels’ machine-audit angle was your deeper-look on May 23; repeat.
• [Rundown — “Google’s AI cracks nine unsolved Erdős problems”]: Fun one-up of OpenAI’s Erdős win (your May 22 conversation capital), but the math-milestone beat is already banked.
• [Rundown — “Starbucks scraps AI inventory tool after nine months”]: Good cautionary tale on deploy-then-discover, but it just reinforces your ESRA adoption thesis (May 24), nothing new to act on.
• [Bagel Bots — “The AI Cost Crisis Is Starting to Hit Big Tech”]: Microsoft/Uber Claude Code cost blowouts already covered (Apr 20, May 23); the cost-discipline point lives in today’s DeepSeek Tier 1.
• [Bagel Bots — “The Prompt That Makes AI Honest”]: Decent facts/assumptions/uncertainty separation prompt, but a single prompt, not a shift — bank it for Ben/AI-Edge due diligence if you want.
• [The Tip — “AI Competitor Weakness Exploiter” / “Specificity Selling”]: Both genuinely useful for MACA copy and UBX verifiability, but they’re craft tips, not news.
• [Practicaly — “sunglasses fitting / use-case library / Zapier templates / Emily Higgins reels”]: Consumer/creator hacks; no project hook.
• [The Information — “Cursor staff meet xAI”, “Twilio’s AI boost”, “1H in 5 charts”, “20 books”]: Subscription-teaser sends; Cursor–xAI already covered, rest off-thesis.
• [Neil Patel — “Last day. Closing tonight.”]: Ubersuggest lifetime-deal promo.

• Sources scanned: 9 newsletter feeds queried (roy.s.mcpherson@gmail.com); 7 returned content (TLDR, The Rundown, The Information, Practicaly, Bagel Bots, The Tip, Neil Patel). Agent AI and a16z silent; second account unavailable this run.
• Items extracted: ~40
• Items surfaced: 8 (3 Tier 1, 2 anxiety-flip, 2 deeper look, 1 conversation capital)
• Items skipped: 13 (covering ~25 underlying items)
• Read time: ~5 min