the boring layer is where the money went

A small glowing data center floats above a dark ocean, connected by one fiber cable branching into light threads, over faint coins and circuits.

Week ending November 14, 2025; archive coverage: November 10.

Everyone wanted to talk about leaked image samples this week. Fine. Meanwhile the actual load-bearing stuff shipped quietly: the scaffolding that makes coding agents useful, the evals that make their scores mean something, the pricing knobs that decide whether you can afford to run them all day, and the trillion-dollar pile of hardware and wet cable the whole thing sits on.

That's the real story, and it's less photogenic than a pretty picture. It's also where you should be looking, because none of the demo magic survives contact with a 50,000-line codebase or a severed fiber in the Red Sea.

the scaffolding is the product now

The most useful thing I read all week wasn't a model release, it was a reference library of Claude Code infrastructure extracted from six months of real TypeScript microservices work. It's honest about being a parts bin, not a demo, which is exactly why I trust it.

The core insight: skills that just sit there waiting for you to remember them are worthless. So there's a UserPromptSubmit hook that reads your prompt and file context, checks a skill-rules.json, and surfaces the relevant skill automatically. Pair that with the 500-line modular pattern, where the main skill file is a navigation index and deep material lives in resource files loaded only when needed, and you get an agent that doesn't drown in its own context.

The catch is it's fitted to one person's monorepo. The Stop hooks reference specific service names, the settings are examples. But the two essential hooks work out of the box, and the practical consequence is real: agent quality is now mostly a scaffolding problem, not a model problem. Copy the hooks, add one skill, move on.

the eval finally got harder and cleaner

Terminal-Bench 2.0 and Harbor is the other half of this. Version 1.0 got popular fast and accumulated flaky tasks that broke when some third-party API changed. So 2.0 is 89 tasks, each run through hours of manual and LLM-assisted validation, with the unstable ones like download-youtube cut.

Here's the part I respect: co-creator Alex Shaw admits SOTA on 2.0 is roughly the same as 1.0 despite the new set being harder, and attributes that to higher task quality rather than pretending the numbers went up. Harbor is the runtime that ran tens of thousands of rollouts to build it, testing any container-installable agent across cloud containers. Codex CLI on GPT-5 leads at 49.6%, which is to say nobody's solving even half. That clustered, sub-50% leaderboard is the honest picture, and I'll take it over any self-graded benchmark.

cheaper knobs for the daily grind

On the practitioner side, OpenAI shipped a few Codex updates: GPT-5-Codex-Mini for roughly 4x more usage at a slight capability cost, 50% higher rate limits for Plus/Business/Edu, and priority processing for Pro/Enterprise. Codex even nudges you toward Mini at 90% of your limit.

This is model selection as a cost/latency dial. Run the small one on easy tasks, save the big one for the hard patches. Unglamorous, but it's the difference between an agent you use and one you ration.

the bill, and the cables under it

All of this rests on spending that's hard to fathom. Tomasz Tunguz's model of OpenAI's $1.15 trillion in hardware commitments across seven vendors implies OpenAI would need to hit roughly $577B in revenue by 2029, about the size of Google today, to make the margins work. His estimates carry a stated 30-50% error bar, so treat it as a shape, not a fact. But the shape is a near-vertical growth curve, and betting careers on it.

And compute is worthless without connectivity. The reminder came from subsea cables: $13B of investment for 2025-2027, nearly double the prior three years, with hyperscalers now owning cables outright. Meta's Aime put it best: without the cables between them, data centers are just expensive warehouses. Add the sabotage worries in the Baltic and around Taiwan and you remember this is physical, fragile stuff.

what I'd do Monday

Steal the two essential Claude Code hooks and wire up auto-activation for one skill you actually use. Adopt Terminal-Bench 2.0 or a Harbor-style containerized harness as your yardstick and stop trusting vendor scores. Put GPT-5-Codex-Mini on your cheap tasks and reserve the big model for real patches. And when someone sells you AGI-by-Tuesday, remember the growth curve it's underwritten by, and that a dragging anchor can take a country offline.

vocabulary inflation

Progressive disclosure — loading a skill's overview first and its deep resource files only when needed, to stay under context limits. Rollouts — individual agent runs against a benchmark; Harbor did tens of thousands to validate Terminal-Bench 2.0.