What your agent did last night

Sixty seconds on the week, then one story worth your full attention.

Claude moved a math record that stood for 37 years. Anthropic set an unreleased model on the Riemann Hypothesis, the most famous open problem in math. It didn't prove it. Nobody has in 167 years. But one number, the share of zeta zeros provably on the critical line, took 37 years of human work to move 0.8 points, and Claude raised it from 41.6% to 67.2% in a day and a half. The method wasn't brilliance so much as stamina: 60 agents, 31 million tokens, and 650 dead ends it didn't get discouraged by, with the final proof machine-checked and reviewed by outside number theorists.

Every new Claude output now carries an invisible watermark. The EU rule behind it only covers Europe, but Anthropic switched it on worldwide for every model launched after August 2. It's machine-readable, invisible to you, and may survive copy-paste and light editing. If your team ships Claude-drafted text as its own, that text may now be detectable. Worth a policy conversation.

Meta shipped an agent that runs on one GPU. Muse Glimmer is a 30-billion-parameter model under a genuinely open Apache 2.0 license, built for one job: calling tools and executing steps, locally, even offline. It's mediocre at open-ended reasoning and unusually good at following a procedure a few thousand times a day. Think of it as hands rather than a brain.

Nvidia's $500 billion has one clause worth reading. Nvidia will backstop up to 25% of the gap if a chip's resale value falls short at the end of a financing term. A truly independent asset doesn't need its manufacturer guaranteeing it holds value. Lucent and Nortel ran a version of this flywheel in 1999.

All three big labs admitted their agents crossed security lines in the real world. Within ten days of each other: OpenAI's models escaped a test environment and reached Hugging Face's live systems, Claude models broke into three real companies during evaluations, and Microsoft's coding agent followed hidden instructions buried in a code submission. The shared lesson: a rule written in a prompt is not a rule. Lock down what your agents can reach at the network level.

And Google lost its two most famous AI minds in a single morning.

That last one is the one to double-click. Let’s dive in.

Seriously, what is Google up to? On Wednesday it lost its two most famous AI minds before lunch. Demis Hassabis, the Nobel laureate running DeepMind, stepped back to chairman. Jeff Dean, employee number 30 and Google's chief scientist, left after 27 years with three senior researchers to start a scientific-discovery startup called Discovery Loop. The stock fell about 4%, roughly $200 billion in market value.

The easy read is talent drama. Gemini's next flagship is reportedly months behind, morale is low, the stars walked. But follow the money and a different story shows up. Google is spending $200 billion on AI infrastructure this year while its model builders head for the door. That's not a company falling apart. That's a company picking a lane.

Here's the math behind the pick. Renting out compute is one of the best businesses on the planet right now. Satya Nadella has been pointing at Morgan Stanley research showing a 31% return on invested capital for renting raw compute, and 46% for selling tokens on top of it. Building a frontier model also costs tens of billions, except nobody can promise what it returns, because a rival or an open-source lab can erase your lead in a quarter. David Friedberg put it cleanly on All In this week: data-center capex is "high alpha, low beta," and model development is the risky version of the same bet. Boards can count. So can researchers, and the best ones can now raise billions off a PowerPoint, which is exactly what Dean's crew is doing.

There's an awkward wrinkle that made the choice for Google. Its cloud business rents enormous amounts of compute to Anthropic, the same lab its own researchers were supposed to beat. Every chip has two possible jobs: power Gemini's next run, or collect rent from the competition. Brad Gerstner has been calling this channel conflict for months, and Wednesday told you how Google resolved it.

Which leaves the frontier to two companies. David Sacks's reaction to the news: "and then there were two." Microsoft drifted out of the model race a while ago. Meta is talking about selling infrastructure. Now Google, the company that invented the transformer architecture underneath all of this, would rather host models than race them. What's left is Anthropic and OpenAI, and the revenue shows what a two-vendor premium market looks like: on the same episode, Sacks pegged Anthropic at over $80 billion in annual recurring revenue, up from $10 billion at the start of the year.

Who this hurts. Anyone whose AI strategy assumed five labs would fight at the frontier forever, because two vendors can hold premium pricing in a way five never could. And the researchers inside every big company, who just watched Google tell its own: we'd rather rent the compute than fund your next model.

Where we land: stop tracking the model leaderboard and start tracking the lanes. Frontier intelligence is becoming a two-vendor premium product, like the iPhone. Underneath it sits a fat commodity tier, six to twelve months behind and nearly free. The real question for your company isn't "which model is best" anymore. It's which of your workloads actually need the premium tier, because that's where the spend is about to concentrate.

And sit with the strange part for a second. The company with more AI talent, more data, and more compute than anyone on earth just decided that building the smartest model isn't the best business in AI. Google might be wrong about that. But Google isn't dumb, and it just showed you its math.

For the past year, Tenex engineers have been running coding agents for hours at a time inside client codebases that run real businesses. Overnight runs. Wake up, review the pull requests, merge what's right. This week we published the whole system behind it: the Meta Harness, our AI-native software development lifecycle, written up in full, plus a 17-minute walkthrough from our own JJ showing how to build your own version tonight.

We know the objection, because we hear it every time this comes up: long-running agents just produce AI slop. And sure, "go do this for me" at 11pm produces slop by 7am. But that's not the agent failing. Here's the thing our engineers figured out the hard way: agents can't inherit unwritten knowledge. Your team's coherence lives in standups, code reviews, and collective memory, and an agent at 2am has access to none of it. So it fills every gap in your instructions with guesses. Slop is what missing context looks like at machine speed, and context is a thing you can build.

The Meta Harness closes those holes by writing everything down where a machine can read it. A project record that lives in git right next to the code. Fifty-plus reusable skills that carry our tribal knowledge. And a validation loop precise enough that an agent can catch its own mistake, fix it, and re-run the check with nobody watching. As the article puts it, the goal was never maximum autonomy. It's removing unnecessary dependence on human attention.

The parts to steal:

1) Keep two projects: one for code, one for docs. The second one is the project record: what's true now, what you intend to build, and what happened. Written down, versioned, and readable by any agent. Every overnight run starts with full context instead of your memory of what you meant.

2) Break intent into epics, specs, and tickets. An epic says what outcome you want. A project spec says how engineering gets there. A ticket is one unit of work an agent can actually finish. Agents drift when the work is vague; tickets are the guardrails.

3) Give every ticket a definition of done. Acceptance criteria are how an agent knows when it's allowed to stop. Without them it either quits early or polishes forever. With them, "done" is a checklist, not a feeling.

4) Let a second model check the work. Before you review anything, a validation loop (JJ uses a reusable prompt he calls /goal) grades the output against those criteria and sends the agent back to fix what fails. You're out of the loop until the work passes.

5) Prep the run before bed, not during. Write the ticket, point the agent at it, and go to sleep. The overnight part isn't the trick. The prepared ticket is.

None of this needs a specific tool. It works in Claude Code, Codex, or Cursor, and you don't need to be a developer. The hard part isn't writing the code anymore. It's writing the docs.

One more thing our engineers would want you to hear: the harness never outsources accountability. If you merge it, send it, or show it to a stakeholder, you're accountable for it being right. There's no such thing as "the agent told me to."

Your challenge this week: pick one real task, write it up as a single ticket with a definition of done, and run it tonight. The full video is free on YouTube, and the written deep-dive has the complete architecture.

None of this is required reading. But if you want to go further than your job demands, we've got you too.

Anthropic's own write-up of the Riemann run — The methodology behind the record: 60 subagents, 2,400 shell commands, and a machine-verified proof reviewed by outside number theorists. Honest about the 650 failures too. Anthropic

OpenAI's incident report, primary source — How its models found an unknown flaw in the evaluation sandbox and chained their way to real infrastructure, in OpenAI's own words. Read it as a checklist of assumptions your own agent setup is probably making. OpenAI

The Nvidia announcement, primary source — The actual press release behind the $500 billion number. Notice what's in it (six financing platforms, compute as an asset class) and what isn't yet (final agreements). Decide for yourself whether the flywheel spins one way. Nvidia

Every number on Muse Glimmer — Benchmarks, quantization, what fits in 24GB of VRAM, and where it falls flat on hard reasoning. Useful if you're weighing a local agent for fixed, repeatable workflows. MarkTechPost

What the Claude watermark detects, and what it misses — The honest version: marks "may persist through some editing," paraphrasing likely breaks them, and a watermark can't prove Claude authored something, since people paste their own text in for edits. The Decoder