The promotion nobody announced

Everyone Who Stayed Just Got a Team

The most misread number this quarter is Salesforce's engineering headcount. It hasn't grown in two years, and the doom take wrote itself: AI froze hiring, the engineers are next. The more useful version is the opposite. The 15,000 people already there got so much more powerful that the company decided it didn't need more of them.

The backdrop is real, to be clear. Tech layoffs hit 142,000 in the first five months of 2026, up 33% from last year, at companies posting records while pouring roughly $700 billion into AI infrastructure. But listen to what Marc Benioff actually said on the May 28 earnings call. Engineering has been flat at around 15,000 for two years, "we're not hiring more engineers," and the reason is "we've been using AI to create more efficiencies for our engineers." That's not 15,000 people doing the same job with fewer coworkers. Each one picked up an amplifier.

Cognition, the maker of the Devin coding agent, just raised over $1 billion at a $26 billion valuation on $492 million of revenue, with Devin shipping production code at NASA, Goldman Sachs, and Mercedes-Benz, growing 50% month over month. The easy story is that the agent replaces the engineer. What's actually happening is that somebody at NASA now directs a fleet of these and ships what used to take a department. The work moved up a level, to the person pointing the machine.

Our read: the line that matters has almost nothing to do with how technical your job is. It's whether you direct the work or wait to be handed it. The person who got promoted this year is the one who treats an agent like a new hire on their team, hands off the grunt work, checks it, and ships. The person sweating is the one still waiting for someone to tell them what an agent is even for. That gap is going to decide careers for the next five years, and your job title barely factors into it.

Even Jensen Huang is making the optimistic case, in a backwards way. He told other CEOs to quit blaming AI for layoffs, calling the narrative "just too lazy." He's right that AI isn't doing the work on its own yet. It still needs a human setting the goal, holding the leash, and catching the mistakes, and that human is worth more now, not less.

The headcount stopped growing because each head got more powerful. That's the promotion nobody sent an email about.

You're Still Using Claude Like a Search Box

So how do you actually become the one holding the leash? It starts somewhere less glamorous than you'd expect: the plain chat window almost everyone treats like a search box. The people getting real leverage out of Claude didn't find a secret feature, they just set the thing up properly, which almost nobody bothers to do.

We regularly train mid-sized teams on how to work with Claude, and what keeps surprising us is how many capable people still live in "chat mode." One question, one answer, repeat. The gap between okay results and scary-good results isn't the model, it's the time almost nobody spends building a foundation that makes every future request better than the last. Set it up once and it compounds, because your preferences, your writing voice, and your tools all stack on top of each other instead of starting from zero every chat.

1. Build the foundation once: In Settings, if you've been using ChatGPT, import your history so Claude starts out knowing how you work. Then have Claude write your global instructions for you. The move most people miss: don't write them yourself, just say "interview me and then draft my instructions." Claude is a better prompter than you are, and this is the highest-leverage 20 minutes you'll spend.
2. Give every prompt a role, context, task, and format: Stop typing "write me an email." Tell Claude who to be ("you're a CMO"), the context ("I run a 10-person agency"), the specific task, and the exact format you want back. Then let it ask you clarifying questions instead of guessing.
3. One task, one chat, and run a few at once: Break the habit of one giant master thread. A fresh chat per task is faster, cheaper on tokens, and sharper. The upside nobody mentions: you can have three or four chats working in parallel, like a small team.
4. Clone your own writing voice into a skill: A skill is just a long, saved prompt you can reuse. Connect your Gmail, then ask Claude to read your sent mail from the last 30 days and build you an "email voice" skill from how you actually write. From then on, "write an email" drafts in your voice, not a robot's.
5. Connect the three tools you live in, then set the trust ladder: Hook up the apps you use all day, Gmail, Drive, Calendar, a notes tool. Start each one on "ask before acting" so Claude checks with you, then graduate the ones you trust to "always allow." Treat it like a new hire who earns the keys over time.

Most people think they're bad at AI. They've just never spent the half hour that turns a generic chatbot into something trained on them. We're beyond Googling and beyond plain chatting now.

Also worth your time

Anthropic shipped Claude Opus 4.8 just 41 days after 4.7

By Anthropic's own measure the new model is about four times less likely to let a flaw in its own code slip by, it ships a Fast mode that runs 2.5x quicker and 3x cheaper, and it adds a research-preview "Dynamic Workflows" that plans a job and farms it out to hundreds of helper agents inside one session. The model gains are real, but the 41-day cadence between major releases is the actual story to watch. (via Axios)

CNN dragged Perplexity into court

CNN filed in the Southern District of New York on May 28, alleging Perplexity scraped more than 17,000 of its stories, photos, and videos, plus a trademark claim that Comet Plus falsely advertised access to CNN's premium content. Perplexity's public defense so far is "you can't copyright facts," which sidesteps the actual complaint. Nine publishers now have active suits, while Time and Gannett quietly took licensing deals instead.

Anthropic filed confidential IPO paperwork

The company submitted a draft S-1 to the SEC on June 1, right after a $65 billion raise that put it at a $965 billion valuation on a revenue run-rate that reportedly jumped from $10 billion to $47 billion in a year. If markets cooperate it's a trillion-dollar debut, and OpenAI and SpaceX are close behind.

A UC Berkeley team showed most AI agent benchmarks can be gamed

Their paper, "How We Broke Top AI Agent Benchmarks," demonstrates agents hitting perfect scores without solving a single real task. One group gamed 890 tasks with a single-character change the benchmark never caught. Next time a vendor quotes you a benchmark number, the right question is which one, and whether anyone's tried to break it.

The Model Context Protocol shipped its biggest revision since launch

The release candidate, locked in late May, goes stateless, adds an Extensions framework, introduces server-rendered sandboxed UIs called "MCP Apps," and adds a Tasks extension for long-running work. If your team builds anything that connects agents to tools, this is the spec to read before it finalizes.

See you next Tuesday. Sven