📅 September 1, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 24
You're in the loop — Tim Cook just handed Apple's keys to a hardware lifer whose very first assignment is fixing the company's AI — and the Siri he'll show off is being rebuilt on a rival's model, Google's Gemini. Meanwhile, half a world away, China's DeepSeek is quietly closing a round that values it at $74 billion, on a straight line toward a 2027 IPO.
Today: a self-audit trick that makes any AI agent check its own work, five fresh tools, a pre-mortem red-team prompt, and the posts lighting up the timeline.
🔁 The Loop
A CEO swap, a $74B raise, and a coding-tool breakup
A leadership handoff and a rebuilt assistant top today's board. Original illustration for InTheLoop.
Apple's new CEO inherits an AI to-do list. John Ternus takes over as Apple CEO today, with Tim Cook shifting to executive chair after 15 years running the company. Ternus, hardware chief since 2021, steps in mid-cycle with one headline problem: a Siri overhaul Apple is now rebuilding on Google's Gemini rather than its own models. His first real test comes September 9, when Apple is expected to unveil a foldable iPhone, new Watches, and the rebuilt assistant — so watch whether a hardware boss can finally make Apple's AI story land. Read more.
DeepSeek is raising at a $74 billion valuation. China's DeepSeek is closing a roughly 50-billion-yuan ($7.4B) round that values it near $74 billion — up from about $50 billion in June — with battery giant CATL and existing backers Monolith and Shixiang joining, per reports. The cash targets model R&D and roughly a gigawatt of new compute, setting up a possible IPO filing by year-end and a 2027 debut on Shanghai's STAR Market. See the details.
OpenAI is cutting Cursor off. OpenAI told coding-tool maker Cursor it intends to end the contract supplying its models, with a proposed cutoff of November 12 — a split that lands as the AI-lab rivalry hardens into open supply-chain warfare. It leaves one of the most-used coding assistants leaning harder on Anthropic, Google, and open weights. Catch up.
Going deeper: why fixing AI is Ternus's defining test, and Apple's own 8-K on the handoff.
🌊 DEEP CURRENT
Everyone deployed agents. Now comes the ROI reckoning.
Weighing the payback: adoption is near-universal, returns are not. Original illustration for InTheLoop.
The gold rush is real. Agents are no longer a pilot-project curiosity: about 79% of companies say they've adopted AI agents somewhere in the org, and 80% of enterprise apps shipped or updated in Q1 embedded at least one — up from 33% back in 2024. On paper, the agentic era already arrived.
But the payoff is patchy. Only 23% report significant ROI from agents (and 29% from generative AI overall). The bright spots are real where teams commit: 41% of agent deployments hit positive payback within a year, the median time-to-value is 5.1 months, and narrow use cases like SDR agents pay back in about 3.4 months.
Only 31% of enterprises have even one agent in production — the gap between "we're using AI" and "it's actually paying off" is where 2026's budgets get won or lost.
The catch nobody advertises. The same surveys show 69% of companies planning layoffs tied to AI — yet Gartner expects 40%+ of agentic-AI projects to be scrapped by 2027 on unclear ROI and weak risk controls. Cutting headcount against tools that a third of buyers can't yet measure is a fragile bet.
The bottom line. Treat agents like hires, not features: pick one workflow with a measurable payback, instrument it, and expand only when the number actually moves. Watch for the first wave of quiet project cancellations this fall as that "cancelled by 2027" forecast starts to bite.
🛠️ The Workbench
Make your AI agent check its own work
Models are far better at catching mistakes than avoiding them. A quick second-pass "self-audit" turns one confident draft into a reviewed one — no extra tools required.
Draft as usual — ask for the deliverable (report, code, email) in one clean pass.
Open a fresh turn and paste the output back in, framed as work the model did not write.
Ask only for a numbered findings list — factual errors, unsupported claims, edge cases — with no rewriting yet.
Have it rate each finding high / medium / low and propose the smallest fix for each.
Now request a corrected version that clears every high-severity item, and diff it against the original.
Sample Prompt: "You are a meticulous, skeptical reviewer. Below is a draft you did not write. List every factual, logical, or edge-case problem as a numbered list, each tagged severity (high/med/low). Do not rewrite anything yet."
🗣️ Overheard
What the timeline's buzzing about
🖥️ Construct Computer: an "AI coworker" that gets its own cloud desktop — browser, terminal, files, and persistent memory — topped Product Hunt with 313 upvotes.
🔓 Session heist: Anthropic is signing users out and wiping saved payment methods after infostealer malware siphoned live Claude login sessions off infected PCs.
💬 Show and tell: the monthly "Ask HN: What are you working on?" thread is packed with solo devs shipping agent side-projects this month.
⚡ Grok goes long: SpaceXAI's Grok 4.6 — 500k context, self-checks its work — is now live in Cursor, Vercel, and Cloudflare.
🎓 Back to school: Anthropic quietly launched Claude Academy, an official hub for getting fluent in Claude.
🔎 Fresh Finds
Five tools worth a look
🎬 Scrollytelling: AI-built immersive, scroll-driven stories and sites with zero design skills.
📣 Oppora AI: automates lead-gen, outreach, and meeting booking for sales teams.
🎨 Shape: an agentic IDE where designers and engineers build in one shared canvas.
📱 bitdrift.ai: agentic mobile observability that watches your app's real-user sessions.
🤝 MeetStream AI: one API to drop AI notetakers and agents into any video meeting.
★ = sponsored placement, if any.
🧪 Prompt Lab
The pre-mortem red team
You are a hostile red-team analyst. Assume the plan below fails badly six months from now. Working backwards, write the 5 most likely causes of that failure, ranked by probability. For each: an early warning sign I could watch for, and one concrete mitigation. Be specific and unsentimental. Plan: [paste your plan]
Prefer a visual? Generate a header for it in a modern-gouache style:
A calm modern gouache illustration of a lighthouse sweeping a foggy harbor at dusk, a warm amber beam cutting through deep-indigo mist, textured paper grain, generous negative space, friendly rounded forms; no text, no letters, no logo, no watermark.
⏪ Rewind
Yesterday’s most-opened link
The enterprise-AI adoption report everyone kept forwarding: why 79% of companies now run agents but most still can't prove the payback. Worth five minutes before your next AI budget meeting.
Stay in the loop — the InTheLoop team