May 10th, 2026

Tier 1 — The Information: per-seat SaaS pricing is cracking under agents (CourseBuilds, MACA, Ben)

The Information ran a Deep Research piece overnight: SAP, Salesforce, ServiceNow are racing to either restrict outside agent access to their platforms or rebuild pricing around AI-driven activity instead of seat counts. ServiceNow’s Action Fabric tollgate (covered May 7) is now part of a pattern, not a one-off. Verdict: research preview / pricing experiments shipping in production. This is the macro signal that validates everything you decided about MACA cost-per-run telemetry, Ben’s CFO budget cap, and CourseBuilds’ embedded annual pricing — none of which are seat-based. Action this week: lift one paragraph of this thesis into the CourseBuilds Aria pitch deck and into the MACA pricing brief. The story is “agent-economics is the new SaaS-economics” and you’re already on the right side.

Tier 1 — Cloudflare 20%, Upwork 25%, Airbnb 60% AI code: the CourseBuilds wedge just got sharper

Stack of CEO admissions in 48 hours: Cloudflare cut 20% citing AI vendor cost surge (covered yesterday), Upwork CEO Hayden Brown announced 25% workforce cut yesterday explicitly framing it as “AI-native restructuring,” Airbnb says AI now writes 60% of new code with the support chatbot resolving 40% of tickets autonomously, Uber publicly slowed hiring on AI productivity gains. Verdict: verified shipped — these are public CEO statements with specific numbers. Every one of these is direct ammo for the Aria conversation with Zaicek. The pitch isn’t “AI is coming” — it’s “your competitors’ competitors already restructured. Here’s the cost of being second.” Update the CourseBuilds pitch deck (~/Reeve/docs/superpowers/specs/2026-04-14-coursebuilds-bespoke-pilot-design.md §wow-moment) with these four data points before next Aria touchpoint.

Tier 1 — Cursor staff meeting with xAI as layoffs and exits mount (tooling continuity risk)

The Information broke last night that Cursor staff have been taking meetings with xAI employees as Cursor lays off and bleeds talent. Verdict: verified reporting (The Information exclusive), early-stage signal not yet a confirmed acqui-hire. This matters because Cursor sits in the dev tooling layer adjacent to Claude Code, which Reeve, Ben, MACA, and CartQuote all depend on. Not an emergency — Claude Code is a separate stack. But if Cursor folds into xAI, the IDE-agent landscape consolidates faster than expected and Codex’s daily-install surge (a16z noted yesterday) suddenly looks structural rather than novelty. Action: nothing today. Note for the Tech Scout idea file — this is the kind of signal that justifies that agent existing.

You priced agent-activity, not seats — months before The Information called it

Ben’s PaperClip integration (~/Developer/PrevailPartners/products/agents/XeroAgent/ben/tools/paperclip_client.py, registered as CFO with $50/mo budget cap, agent-id 50113ed1) charges by task completion + cost reporting, not by user. MACA’s api/lib/costs.ts per-run cost logger and public/cost-dashboard.html were built on the same instinct. CourseBuilds’ $50-120K/year embedded pricing in the spec doc is annual-engagement, not per-seat. The Information’s tollgate piece arrives today asking the question you already answered in March: when agents do the work of multiple seats, what’s the unit of value? You’ve been shipping the answer for two months. When Zaicek asks how Aria should price internal AI workflows, you have the receipts.

Anthropic’s free 24-min prompting workshop — actually watch this one

Prompting 101 | Code w/ Claude — flagged in Practicaly yesterday but burying the lede. It covers memory, automation, systems-thinking prompts. You’re about to invoke superpowers:writing-plans to convert the UBX South Bank sale spec to an executable plan. You’re scoping the CourseBuilds Phase 0 wow-artefact library. The Always-On Reeve Phase 2 build needs sharper headless prompts. 24 minutes from Anthropic’s own team. Watch it on the train, take notes against ~/Reeve/reeve-headless.md and the CourseBuilds wow-artefact templates. Specific angle for you: the “systems-thinking prompts” section is the missing piece between Roy-as-prompter and Reeve-as-autonomous.

Six-hour Codex /goal run that survived a five-hour pause

tectontide.com/en/blog/codex-goal-six-hour-run — Codex’s persisted goals survive terminal restarts, laptop sleeps, and multi-hour pauses without re-prompting, with a developer message injected on resume. This is the exact pattern Always-On Reeve Phase 2 needs (persistent Telegram listener + iMessage config + content processing pipeline). 30 minutes of read-and-port: which of these mechanics translate cleanly to PaperClip + claude_local? The injection-on-resume vs heartbeat-checklist tradeoff (~/Reeve/HEARTBEAT.md) is worth a paragraph in the next Reeve research doc.

“The interesting thing this week isn’t that Cloudflare cut 20% or Upwork cut 25% — it’s that Airbnb’s CEO is publicly saying 60% of new code is now AI-written and the support chatbot resolves 40% of tickets autonomously. Meanwhile The Information’s running pieces on SAP and Salesforce trying to pivot off seat-based pricing because agents break the model. The question for any operator right now isn’t ‘will AI affect my P&L,’ it’s ‘will my pricing structure survive when one of my customers replaces five seats with one agent.’”

Use case: Drop into the Aria conversation with Zaicek the next time the room turns to “what’s actually happening with AI.” Signals: you read primary sources, you connect macro CEO admissions to specific operational decisions, and you’ve already answered the pricing question for your own products. Also lands in the RT R53597 interview if the hiring manager asks about enterprise AI shifts.

SoftBank scaled back its OpenAI margin loan as lenders questioned the $852B valuation

Quietly reported by Zerohedge yesterday and barely picked up by the major outlets — SoftBank pulled back a planned margin loan backed by its OpenAI stake after lenders balked at the reliability of OpenAI’s $852B valuation. Combine with The Information’s overnight reporting on OpenAI’s Broadcom $18B chip-deal financing snag and Musk’s legal expert pegging the OpenAI Foundation at $200B in court testimony. First-mover read: if private AI funding cracks even slightly, the “AI tax” you’re factoring into MACA / CourseBuilds / Ben pricing assumptions (e.g. Sonnet/Opus/GPT API rates trending up forever) might be wrong — there’s a non-trivial path where compute pricing gets more competitive in late 2026 as lenders force discipline. Act/queue/drop: queue. Don’t change pricing assumptions today, but add a note in memory/project_agent-architecture.md to revisit API cost forecasts in Q3 if private funding stress signals continue. This isn’t a project you chase — it’s a forecast variable you watch.

• [TLDR — “Codex in Chrome”]: covered May 8, no material update.
• [TLDR — “Inside Chinese AI labs”]: interesting essay, not actionable for any current project.
• [TLDR — “Long AI Short AGI”]: thesis piece, no decisions hinge on it.
• [TLDR — “Notes from inside China’s AI labs”]: cultural analysis, parked.
• [TLDR — “AlphaEvolve impact”]: DeepMind product update, not in your stack.
• [TLDR — “Good QC for RL data”]: vendor-side concern, you’re not selling RL data.
• [TLDR — “Prime-RL Fast Ask”]: deep ML engineering, not the level you need.
• [TLDR — “Meta IKBO recsys”]: kernel optimisation, irrelevant to your stack.
• [TLDR — “ds4.c DeepSeek V4 Flash”]: local inference engine, alpha, parked.
• [TLDR — “GitHub token efficiency”]: useful but you’re not running GitHub agentic workflows at cost-relevant volume.
• [TLDR — “Anthropic NLA”]: alignment research, not pitch material.
• [TLDR — “Trusted Contact in ChatGPT”]: safety feature, no project relevance.
• [TLDR — “Perplexity Personal Computer Mac”]: covered, monitor only.
• [TLDR — “DeepMind + EVE Online”]: cute, not actionable.
• [Rundown — “Google Health / Fitbit Air”]: covered yesterday, no update.
• [Rundown — “Anthropic Institute self-improving agenda”]: covered yesterday.
• [Rundown — “Spotify Personal Podcasts”]: consumer feature, irrelevant.
• [Rundown — “Scale AI $500M Pentagon”]: covered yesterday.
• [Rundown — “Mozilla Firefox + Claude Mythos”]: nice anecdote, not actionable.
• [Practicaly — “GPT-5.5 Instant in Microsoft 365 Copilot”]: covered yesterday.
• [Practicaly — “Open Design Claude alternative”]: covered yesterday.
• [Practicaly — “ChatGPT Pro 10x usage to May 31”]: covered yesterday.
• [a16z Charts — “Other Income / Slop Surplus / B2B vendor mix”]: covered yesterday in deeper-look slot.
• [Bagelbots — “Anti-Fluff Filter prompt”]: copy if useful, not pitch material.
• [Bagelbots — “Princeton 3D-MIND living brain cells”]: science, decade-out, skip.
• [Bagelbots — “Scientists found 10K exoplanets with AI”]: cool, not actionable.
• [Bagelbots — “Nvidia $2.1B IREN Texas data centers”]: infra, monitor only.
• [Bagelbots — “Anti-AI data center backlash”]: ESG signal, parked.
• [The Tip — “Internal Process Documenter prompt”]: generic prompt, you have better ones.
• [The Tip — “Symbiotic Brand Parasites 2035”]: speculative essay, covered yesterday.
• [The Information — “Polymarket homecoming shaky”]: covered yesterday.
• [The Information — “Broadcom $18B financing snag”]: covered yesterday.
• [The Information — “Anthropic $200B Google commitment”]: covered May 7.

• Sources scanned: 9 (TLDR AI, The Rundown, Practicaly, a16z, The Tip, Bagelbots x2, The Information x2)
• Items extracted: ~50
• Items surfaced: 8 (3 Tier 1 + 1 anxiety-flip + 2 deeper-look + 1 conversation capital + 1 first-mover)
• Items skipped: 32
• Read time: ~7 minutes