For two issues straight the story was that agents now write the code, and the economics finally close. True, and settled.

The interesting part is what got exposed underneath. VentureBeat put a name on it, NVIDIA paid $400M for the layer that sits above it, DeepSeek raised $7.4B to give its models away, and Anthropic, of all companies, asked the industry to find a brake pedal.

The pattern is the same in all four: the machine can do the work, and the hard problem is now everything around the work.

VentureBeat made the case this week that agents have closed the gap on code generation, and that writing code was never the part that slowed teams down. Defining the right requirements, wiring into messy systems, and keeping software alive in production always were. Agents compress execution. They do not compress ambiguity or accountability.

The positional shift is that human review just became the constraint instead of typing. When a hundred subagents ship code faster than anyone can read it, the engineer who reviews, architects, and owns the call is worth more, and the one who only produces lines is worth less. The piece names the new failure mode plainly: teams lose the context they need to catch the agent's mistakes.

So run the audit this week. Find where your delivery stalls, and it will not be in the writing. Then move your strongest people from producing output to judging it, and put a real approval gate in front of anything an agent can ship to production. The skill to grow is taste, not throughput.

Usage-based pricing is transforming B2B revenue—but finance teams are struggling to keep up. Join Tabs and PwC on June 10th for a live breakdown of what it takes to scale modern pricing models. Save your spot now.

Count Me In

On June 3 NVIDIA bought Kumo AI for more than $400M, a startup co-founded by Slovenian Stanford professor Jure Leskovec and North Macedonian engineer Vana Josifovski, the former CTO of Airbnb and Pinterest. Kumo does not write code. It builds models that predict outcomes straight from a company's own data using graph neural networks, cutting manual data prep by up to 95%, with Reddit, DoorDash, and Sainsbury's already on it.

Read what NVIDIA chose to buy. Not another coding agent, but the layer that decides what to do with data, the judgment tier the code-generation story keeps pointing at. And it paid that price for two scientists the region produced and exported years ago, the same diaspora pattern that turned Fonoa's Croatian founders into acquirers last month. The value is migrating off the part agents commoditized and onto the part they have not.

For builders in the region, the lesson is where to aim. The wrapper around a chatbot is a weekend project now. The model that turns a messy enterprise database into a reliable prediction is what a chip giant pays nine figures for. If you have the math talent here, and the region does, build above the code, not beside it.

DeepSeek is closing its first-ever funding round, around $7.4B at a valuation up to $59B, with Tencent and battery giant CATL the largest backers and founder Liang Wenfeng putting in roughly 40% himself. The stated plan is to keep the models open-weight, free to download, on the way to AGI.

Read what the money is for. The biggest maiden raise in Chinese tech history is being assembled to give the core capability away. When a $59B company's strategy is open weights, the model stops being the moat and turns into the loss leader. Capability is racing toward zero price, which is the same message the rest of the week sends from the other end: the frontier model is becoming the cheap part.

So stop pinning your product to access to one particular model, because the frontier is turning into something you rent by the token from whoever is cheapest this quarter. Build the part that survives the model going free: your proprietary data, your judgment layer, your distribution. Watch whether Western labs can hold premium pricing once an open competitor ships the same benchmark at no cost.

On June 4 Anthropic called for a coordinated global pause framework for the most powerful AI, warning that models are nearing "recursive self-improvement," the point where a system can design and train its own successor faster than regulators or safety researchers can keep up. The company was careful about it: not a halt today, but an agreed brake the industry could pull together if the curve bends.

Read who is asking. The lab whose own agent already writes more than 80% of its code, with engineers shipping several times as much per quarter as they did in 2024, is the one warning the wheel is getting loose. That is the whole week in one move: the capability is not in doubt, the control is. Skeptics note the timing, weeks before a near-trillion-dollar listing, and that a pause means nothing unless rivals sign the same page.

Watch whether any competitor echoes it, because a solo brake does nothing and everyone in the room knows it. The read for operators is blunter than the headline: the people closest to the capability are telling you the governance is not ready. Put the human approval gate, the spend cap, and the kill switch on your own agents now, while it is still your call and not a framework's.

Five tools to install or test this week.

Productivity: Notion turned your workspace into a home for agents. Notion opened its Developer Platform with an External Agents API, letting Claude Code, Cursor, and Codex act as visible collaborators inside your docs and databases instead of off in a separate tab. If your team already lives in Notion, wire one agent in this week and watch where it saves time.

Design: Canva research-to-deck lands inside Perplexity. Canva shipped a connector for Perplexity Computer that turns research, briefs, and notes into editable, on-brand presentations and social assets across 11 languages. If you build decks from scratch every week, run a brief through it and edit the output rather than starting on a blank page.

Cost: Torii shows you the AI bill before it explodes. After the week's invoice horror stories, Torii's AI Dashboard breaks token and seat spend down by employee and model, forecasts overages before the invoice lands, and flags overlapping subscriptions across Copilot, Claude, and Gemini. Point it at your stack before you approve the next agent rollout.

Dev: Claude Code added Dynamic Workflows and Opus 4.8. Claude Code now runs Opus 4.8 with Dynamic Workflows, letting you save and rerun multi-step agent routines instead of re-prompting each time. Move one repetitive task you do by hand into a saved workflow and measure the time back.

Dev: Codex can now run for days without you. OpenAI's Codex added multi-day automations that pick up across sessions, with GPT-5.5 plus Codex now generally available on Amazon Bedrock for teams that want AWS governance around it. Hand it a long migration and review the result, not the keystrokes.

Dev: Codex can now drive the iOS Simulator itself. OpenAI's Build iOS Apps plugin for Codex builds and launches your app in an embedded simulator, opens SwiftUI previews, and hot-reloads edits without leaving the tool, capturing its own screenshots to check the result. If you ship on iOS, hand it a small SwiftUI change and watch it test the output on screen, not just write the diff.