June 17th, 2026

Tier 1 — AT&T tokenminimizes, KPMG hallucinates, the enterprise mood turns

The Information confirms AT&T is throttling employee Copilot access, Meta is capping Anthropic spend, Uber and Walmart have capped coding-agent budgets, Amazon scrapped its AI-use leaderboard, and KPMG had to pull its October enterprise-AI report after Claude hallucinated fake UBS and Swiss Federal Railways case studies inside it. Verified shipped — these are real CIO decisions logged in named companies. Microsoft Copilot Cowork pricing has now mirrored Anthropic’s per-seat-plus-consumption model, locking the pattern in. Action: stop pitching MACA and CourseBuilds against the unlimited-token mental model; pitch them against the cap. Roy’s Apr 20 covered-story on Uber’s Claude Code budget blowout is now the canonical cautionary tale — use it.

Tier 1 — Factory 2.0 reframes coding agents as “software factories” ($1.5B, Nvidia-backed)

Factory shipped 2.0 and pivoted the language from “coding agents” to “software factories” — engineers now “build the factories that build the software.” Verified shipped (TLDR 2026-06-16, $1.5B valuation confirmed in The Information). They also launched a model router, which is the same architectural pattern Ben uses in ~/Developer/PrevailPartners/products/agents/XeroAgent/ to route between Claude variants. The bigger payload for Roy: this is the exact Prevail Partners narrative — Roy isn’t selling agents, he’s selling the factory that builds them (Aria lease abstractor today, Ben for bookkeeping, MACA for ad ops, ESRA for shift reports). Factory at $1.5B externalises that pitch. Action: lift the phrase “build the factory, not the artefact” into the CourseBuilds Tier 2 embedded-engagement pitch this week.

Tier 1 — Free Fable letter + Anthropic class action: stack risk rises but doesn’t shift

100+ cybersecurity leaders (Adobe, Zoom, Sophos, Vercel, Veracode, Nvidia, Stanford HAI) signed the freefable.org open letter calling the Fable 5 export-control shutdown “more about politics than protection.” Alex Stamos went on record saying the jailbreak produced a proof-of-concept defenders need to patch. Separately, Anthropic was sued in federal court Friday for overselling the $200/mo Claude plan — actual usage allegedly trails advertised limits. Verified shipped (open letter live, suit filed per WSJ). Combined message: Anthropic momentum is intact but two flank pressures are real, and Roy’s $200/mo plan is now legally contested. Action: no stack changes — Claude Code Routines + Opus 4.8 are still load-bearing for Ben and Always-On Reeve. But log usage tightly this month; if the suit gets traction, expect either a refund window or a plan restructure that affects Reeve’s headless budget.

MACA shipped per-run cost tracking before tokenminimizing was a word

PR #10 (MACA, merged) wired per-run cost tracking into api/lib/costs.ts with a dashboard at public/cost-dashboard.html and photo-pipeline costs serialised to scripts/photo-costs.json. The Information just made that the table stakes question every CIO is asking. While Microsoft and Anthropic are scrambling to publish usage governance docs, MACA already exposes per-ad unit economics as a first-class surface — that’s the exact buyer-facing artefact Aria and any UBX-Australia operator will demand once they look past the creative. The next pipeline run is now also a sales asset, not just a dev metric. The work was finished weeks ago; the market just caught up.

Faros AI 22,000-developer study: AI code review broke the review (TLDR 2026-06-16)

The cited numbers — code churn up 861%, per-developer defect rate up from 9% to 54%, review duration up 441%, zero-review merges up 31% — are exactly the failure mode CartQuote and MACA are heading into as both codebases pass the “AI writes most of the diff” line. GitClear’s 4x raw output for ~12% delivered-value gain is the more brutal stat. 30-minute angle for Roy: read the study summary, then decide whether to seed ~/.claude/skills/launch-strategist/ with a “AI-authored PR review heuristic” (or extend CodeRabbit config rather than build new). This is the empirical case to finally activate the Custom Code Review Skill that’s been parked.

Devin’s $10M-per-customer guarantee (TLDR 2026-06-15)

Cognition is now guaranteeing more dollar-value engineering output than the contract costs. That pricing structure — outcome-bonded, not seat-bonded — is the model CourseBuilds Tier 2 embedded engagement should mirror: “you spend $50-120K/year; we guarantee X automations live by month 6 or refund.” 30 minutes to read Cognition’s pricing page and lift one paragraph into the CourseBuilds spec at ~/Reeve/docs/superpowers/specs/2026-04-14-coursebuilds-bespoke-pilot-design.md.

“Aaron Levie called it last week — ‘We never celebrated tokenmaxxing. We never had leaderboards, so we didn’t get ahead of our skis on incentivising the wrong thing.’ AT&T just throttled employee Copilot, Meta capped its Anthropic spend, KPMG had to pull an October enterprise-AI report because Claude hallucinated UBS case studies into it. Tokenminimizing is the named pattern now, and the companies that built cost ceilings before the budget conversation hit are the only ones still shipping.”

Use case: Drops cleanly into any Aria or Rio Tinto conversation where the other side is worried about AI runaway spend — and lets Roy reposition Prevail’s MACA cost dashboard and Ben’s PaperClip budget enforcement as the answer, not engineering hygiene.

Sakana Marlin: autonomous research agent in beta with 8-hour single runs (~300 expert pilots)

TLDR + The Rundown both surface Sakana Marlin — drop a topic, walk away, come back to a strategy report and a slide deck. Flexible pricing, ~300 experts in beta. This is the exact shape of the Idea Validation Pipeline Phase 1 skill Roy has at ~/Reeve/ideas/2026-04-01-idea-validation-pipeline.md — except Sakana shipped first and is going market-wide. Wingman call: don’t act now, but don’t ignore. If Marlin’s beta opens, queue Roy for it inside the week — it’s faster to evaluate someone else’s $X/month research agent than to keep deferring the Phase 1 build. If the Idea Validation Pipeline ever gets built post-MACA, position it as “Marlin for Roy’s specific brief library + voice notes + RT operational data” rather than a generic research agent. The differentiator is the corpus, not the loop.

• [TLDR — “GLM-5.2 1M context, MIT-licensed”]: open-source frontier model, not a Roy stack change today.
• [TLDR — “Kimi K2.7 Code 1T-param MoE agentic coder”]: same — interesting OSS coding model, no current pull.
• [TLDR — “Codex Mobile”]: OpenAI’s mobile control surface, doesn’t displace Claude Code Routines workflow.
• [TLDR — “AWS WAF AI traffic monetization”]: relevant only if Roy starts hosting content scrapers want.
• [TLDR — “Should you post-train your own model?”]: theoretical for now, Roy doesn’t have a power-law training use case.
• [TLDR — “Meta AI Mode in Facebook search”]: consumer, not the MACA Meta Ads API surface.
• [TLDR — “Anthropic vs OpenAI context strategies”]: interesting Zvi analysis but not actionable for Ben’s architecture today.
• [TLDR — “Ramp SWE-Bench private benchmark”]: benchmarking news, no immediate read-across.
• [Rundown — “Anthropic Fable shutdown lead”]: covered yesterday by The Rundown — Free Fable letter is the new angle, already surfaced.
• [Rundown — “Salesforce buys Fin (Intercom) $3.6B for Agentforce”]: agentic CRM consolidation noise, no action.
• [Rundown — “Robinhood Agentic Trading via MCP”]: novel MCP use case, but consumer finance — not a Prevail wedge.
• [Rundown — “Cartesia Sonic-3.5 + Ink-2 voice/transcription #1”]: voice stack leader change, but Reeve’s voice pipeline isn’t the bottleneck.
• [Rundown — “Reader Slackbot → Salesforce contact workflow”]: nice “no-code” story, no project match.
• [The Information — “DeepSeek $7B funding”]: macro deal, no Roy decision attached.
• [The Information — “Qualcomm/Tenstorrent talks”]: chip M&A, skip.
• [The Information — “Polymarket / Kalshi fraud rings”]: not AI-stack relevant.
• [The Information — “Nvidia server marketplace $100M”]: infra, skip.
• [The Information — “Anthropic discusses possible accord with Trump admin”]: politics, watch but no action.
• [The Information — “AI consultants, coding agents and rising AI costs”]: editor’s recap of the AT&T story already surfaced.
• [The Information — “Inside Broadcom”]: chip strategy, skip.
• [The Information — “Inside Anthropic’s rise / partner tensions”]: backgrounder, watch — but no action.
• [thetip.ai — “Apple starts over on Siri, brings Google in”]: Siri news, not a Prevail wedge.
• [Practicaly — “Dopamine shopping sites + Opus 4.8 video edit”]: consumer + demo, no action.
• [Practicaly — “Everyone’s rebuilding the banned model”]: Fable narrative noise.
• [BagelBots — “What should you build first?”]: prompt-marketing.
• [BagelBots — “Prompt that creates 5 posts from 1 idea”]: prompt-marketing.
• [Neil Patel — “Manage your reputation on AI”]: AEO repackage, already a covered theme.
• [Neil Patel — “Halfway through the anniversary promo”]: pure promo.
• [a16z — “The a16z CFO Community”]: community launch.
• [a16z — “SpaceX & the Sentient Sun”]: vibes essay.

• Sources scanned: 9 newsletters (TLDR AI, The Rundown, The Information, Practicaly, Neil Patel, a16z, BagelBots, thetip.ai, AgentAI not received)
• Items extracted: ~45
• Items surfaced: 8 (1 PAY ATTENTION, 3 Tier 1, 1 anxiety-flip, 2 deeper-look, 1 first-mover)
• Items skipped: 31
• Read time: ~8 minutes at 250 wpm