AI So Whats
Edition

Saturday, August 29, 2026

8 issues 23h window ~7 min read

The agent stack is consolidating around one routing agent, not orchestration fleets

2 sources

Every's Mike Taylor admits he expected executives to be managing many agents, but the dominant pattern turned out to be a single agent that routes tasks. Anthropic's own shipping log supports it from the other direction: the week's headline features are session continuity across terminal, desktop and phone, plus persistent subagent memory — making one session follow you everywhere rather than multiplying sessions. The orchestration layer vendors are selling is being absorbed into the client.

SaaS is being re-priced on whether AI is a cost or a moat, and the market can't decide

2 sources

The Information's enterprise roundup has Salesforce and Datadog marked up within weeks of looking shaky, Workday up 18% on Silver Lake talks, and Klarna down 20% and staying there. Its Notion piece is the same question at company scale: a 30% headcount increase into AI products, against stalling mobile usage and a conspicuous refusal to raise. Both issues argue the AI-era sort in software is unresolved, and that take-private bids are currently doing the price discovery.

Everyone is arguing AI adoption is slower and more mundane than the capex implies

3 sources

Thompson's breaker's-advantage piece concludes AI's takeover of the economy will take longer than people think because incumbents' incentives punish breaking things. Every's consultants say the best early wins are boring and frequent — meeting prep, call summaries, inbox triage — and that being two months behind the frontier costs nothing. McKinsey adds the consumer side: trust in generative AI recommendations below 40 percent, and $5.1B into humanoids that will be 2% of physical AI's value. Three different vantage points converging on deflated near-term expectations.

This week in Claude Code: /resume on desktop, start sessions from your phone, and more

So what

Your engineers can now run Claude Code on locked-down shared machines and see per-task token spend, which removes two standing objections to broader rollout.

Anthropic's weekly shipping log is mostly about making Claude Code a multi-surface, multi-session product: terminal sessions now resume on desktop, and you can kick off a run on any machine from your phone.

  • /resume pulls a closed terminal session into the desktop app; any machine running claude rc appears as a device card in the mobile Code tab
  • New --restricted mode for shared machines and eval harnesses: no command execution, no web fetch, file tools confined to the working directory, user and project settings ignored
  • Cost controls arrive in bulk: a /claude-api cost-optimize skill, per-task token rows in /usage for /loop, and promptCacheTtl extending the cache from 5 to 60 minutes
  • Subagents get persistent memory via memory: project|user|local frontmatter, writing to .claude/agent-memory/
  • CLI binary now zstd-compressed — 75 MB downloads instead of 340 MB, and 40–70 MB less memory per session
ai agents ×3 agent memory claude code developer tooling Open in Gmail →

Not Every SaaS Stock Is Under Pressure. Here's Who's Gaining Ground.

So what

Take-private bids at Workday and Datadog reset the floor on SaaS multiples — your public comps for private marks just got more generous.

A snapshot of enterprise software being re-priced day by day rather than settling into a stable AI winners-and-losers story, with two names flipping direction inside weeks.

  • Salesforce lifted annual revenue guidance and the stock jumped; it also pulled back a second OpenAI sales exec within 24 hours
  • Datadog +5% on a report Palo Alto Networks explored acquiring it; Nikesh Arora's broader deal hunt is reshaping the security layer
  • Workday +18% on reported Silver Lake takeover talks, feeding hope the SaaS sell-off has bottomed
  • Klarna −20% on a cut outlook and has stayed there — the only name where pressure hasn't eased
  • ClickHouse recurring revenue passed $350M on OpenAI and agent usage; DOJ is probing a16z board seats at Databricks and Fivetran
enterprise software ×2 saas repricing ×2 m&a public markets Open in Gmail →

Pacsun’s Gen Z playbook: Listen, learn, and let go

So what

Skip the retail interview; note the humanoid figure — $5.1B chasing 2% of the physical-AI value pool is a mispricing worth acting on.

A Pacsun CEO interview dressed as a retail strategy note: Brie Olson's pitch is that customer advisory councils, surveys and community forums are what turn a mall brand into a youth-culture brand.

  • Pacsun runs annual target-audience surveys, all-customer advisory councils, and customer forums simultaneously — the stated goal is "a coveted youth culture brand"
  • Olson's framing is listen, learn, and let go: ceding product and brand direction to the Gen Z audience rather than merchandising at it
  • No numbers attached to the strategy in the excerpt; it's a podcast transcript, not research
  • The one hard datapoint sits in the sidebar: $5.1B invested in humanoid robots since 2024, yet humanoids are projected at just 2% of physical AI's value by 2045
consumer ×2 physical ai ×2 mckinsey ×2 retail Open in Gmail →

$52m/yr plaid for crypto payments

So what

Mesh at 39x forward against Airwallex at 7.3x is the number to defend or short — routing leverage has to justify a 5x premium.

Sacra's thesis is that Mesh is the Plaid of crypto payments — one abstraction over 300+ wallets and exchanges, settling to a single stablecoin or fiat currency for the merchant.

  • Estimated $52M annualized revenue in June 2026, up from $31M at end-2025 and ~$1M in January 2024; $2B valuation as of the July 2026 Binance round
  • That's a ~39x forward multiple against Kraken at ~8x (~$2.5B annualized, $20B valuation) and Airwallex at ~7.3x (~$1.5B annualized, $11B)
  • Distribution is B2B2C through processors and wallets, not direct merchant competition with Stripe or PayPal — it powers Pay with Crypto for PayPal and Shift4 across 10M+ merchants
  • The Mesh Alliance Program, launched June 2026, turns five years of hand-built integrations into a marketplace where partners bid for routed volume
  • Kraken paid $600M for stablecoin card and orchestration provider Reap, a comp for the same layer
crypto payments fintech revenue multiples stablecoins Open in Gmail →

Consumers don’t trust AI advice. They turn to it anyway.

So what

Stop treating consumer AI distrust as an adoption blocker in your portfolio's commerce decks — usage is decoupled from trust and the funnel is already moving.

McKinsey's consumer research finds trust in generative AI product recommendations sits below 40 percent — and shoppers use it anyway, which is the more interesting half of the finding.

  • Danielle Bozarth reports "significant skepticism around algorithm-driven channels like social media, and in particular generative AI," with trust under 40%
  • The gap between stated distrust and actual usage is what's reshaping the path to purchase, not trust itself
  • Framed as a brand problem: what companies must do differently as AI intermediates discovery
  • Same event promo attached: $5.1B into humanoid robots since 2024 versus a projected 2% share of physical AI value by 2045
consumer ×2 ai adoption ×2 physical ai ×2 mckinsey ×2 Open in Gmail →

Internet Hype and Real World Change (This Week in Stratechery)

So what

The breaker's-advantage frame gives you a defensible reason to underwrite longer AI adoption timelines without sounding like a skeptic to your LPs.

The week's throughline is that attackers have a structural advantage over defenders — and Thompson extends that asymmetry from agentic cybersecurity to why startups beat incumbents and why AI's economic takeover will be slower than expected.

  • Autonomy and Innovation argues from the Hugging Face Incident that agents favor offense: defenders' first job is not breaking things, while breaking things is the attacker's whole point
  • Netflix is weighing selling access to rival streaming services, reversing a decade of standalone insistence; Sharp Tech's read is Hollywood stayed irrational longer than Netflix stayed patient
  • Andrew Sharp bets the under on the data center backlash — now that tech people are worried, he calls it a coming non-issue
  • Apple's Mini and Studio updates plus OpenAI's Jalapeño hardware both read as pressure on Nvidia; the week's video is Nvidia's Risky Business
  • Sharp China: Unitree stock down 45%
ai agents ×3 netflix nvidia stratechery Open in Gmail →

33 Questions Executives Ask About AI—Answered

So what

Kill any portfolio company's plan for a central AI knowledge hub or all-hands AI training — fund three champions and a shared skill repo instead.

Every's consulting practice publishes the 33 questions 400 executives actually asked, and the answers are consistently anti-infrastructure: raise the ceiling with a few fluent people, don't build a knowledge hub, don't chase tools.

  • Core claim: it's easier for one AI-fluent person to build 10 skills than to get 10 people to build one each — fund champions, not blanket training
  • Pick one platform and stay; anything genuinely good gets copied across labs in eight to twelve weeks, so being a month or two behind the frontier costs nothing
  • Centralized AI knowledge hubs don't work — documents go stale on publication and maintenance is resource-intensive with little payoff
  • Mike Taylor's reversal: he expected everyone managing many agents, but the dominant pattern is one agent that routes tasks; and writing turned out harder than code
  • Governance answer: enterprise Claude activity logs give regulators more explainability, not less; tier scrutiny by whether the tool touches databases
ai adoption ×2 enterprise ai ×2 ai agents ×3 ai governance Open in Gmail →

Can Notion Escape Airtable's Dismal Fate?

So what

Notion declining to raise into an open market is the tell — read stalled mobile usage as the real signal before you mark any workspace-software position.

Notion is betting the company on AI, and the piece frames it as an adapt-or-die test for a 13-year-old pre-AI software brand with Airtable's fate as the cautionary comp.

  • Headcount up roughly 30% this year, with new roles concentrated in building and selling AI products
  • Mobile app usage is stalling per Sensor Tower, and Notion hasn't raised in a market where capital is free-flowing — an unflattering signal
  • New features: custom agents for repetitive tasks, meeting transcription, database automation, and search across Slack and Google Drive from inside Notion
  • Ivan Zhao's design reputation and surging revenue are the bull case; frontier-lab products eating the workspace category are the bear case
enterprise software ×2 saas repricing ×2 enterprise ai ×2 notion Open in Gmail →