You can’t improve what you can’t see

I have a confession. (Long-time readers may notice I've even broken voice, choosing "I" over "we" lest I implicate Claire in my dirty little secret.) I hate, despise even, the software factory analogy. (Even more, perhaps, than I hate that I'm no longer allowed to use em dashes in my writing.)

Look, I get it. We're trying to convey that something that was once bespoke and handcrafted has become largely commoditized by technology. And I'm certainly no expert in manufactured goods. But I'm pretty sure there aren't many factories out there that are designed to create infinitely arbitrary products or parts, and in fact will likely never create the same part twice. It just feels like a bad analogy to me.

But we all seem to have agreed to call systems of software creation and delivery "factories" now that AI is involved, and I do my best not to be overly stubborn or petty, so we'll just call them factories here as well. I'll take solace in the fact that the collective we seem to have agreed that dark factories are not actually the near-term goal, and "lights-on" factories with humans strategically kept in the loop are what we should actually be building.

And factories are having a moment. There's certainly some amount of bending the term to mean whatever is convenient, and also: teams are undoubtedly putting similar pieces together to create high-throughput systems for creating, deploying, and monitoring software with LLMs at the center. If you take a moment to consider where we're at compared to 12 or 18 months ago, it's really quite something.

We have a few articles this week to help demystify the "software factory" for those of you who aren't there yet, or guide you toward success if you are. At its most basic, I think the test is this: can you see what the system is doing well enough to improve it, and use that feedback to reliably deliver software without bottlenecking on human toil (bottlenecking on human innovation and judgment is, as far as I can tell, still acceptable).

Here's what we're reading this week.

Closing the loop with self-improving cloud software factories

Warp literally sells a product called Factories, so do what you will with that, but CEO Zach Lloyd consistently writes with clarity and restraint on their viewpoint of what makes for a good software factory. If you're thinking about building your own factory, looking to buy factory-like things from a vendor, or just trying to figure out what a software factory is, this is a great primer with an opinionated but balanced point of view.

Software factories are automation loops around the SDLC, comprised of agents that triage, spec, implement, verify, review, monitor, etc. Done right, these factories allow for true measurement, improvement and automation over time, and can help prove that you are doing agentic engineering the right way.

The Warp team is big on measurement as a core piece of the factory puzzle, and we agree. One of the first questions we ask clients, whether they're piecing together factory-like things or still struggling with AI adoption and fluency, is: what are you measuring? If you're not measuring, you're flying blind. If you don't know what to measure, you don't know what you're trying to inflect.

Running a Software Factory Efficiently at Uber Scale

Speaking of measuring things. From the company that famously burned through its entire 2026 AI budget by mid-April, we now have a concrete example of how to move past tokenmaxxing leaderboards to sustainable practices. Uber got there by leaning on its strengths and treating the cost problem like any other engineering problem. Rather than do an about-face and tell everyone to stop using AI, they broke it down into things they wanted to continue to grow (users of AI, sessions per user) and things they wanted to shrink or optimize (turns per session, requests per turn, tokens per request).

From February to Aug 2026, weekly active users across all agentic offerings across all our employees (engineers & non-engineers) grew 7x, and weekly agentic requests grew 9.4x. Meanwhile, our total AI spend has relatively stabilized since April due to optimizations across the board.

You can't optimize if you don't measure. Uber lays out a concrete list of things they measure and walks through how they went about optimizing them. It's a long post, but chock-full of amazing detail, so it's worth your time if you're at all interested (and you should be).

Ramp's coding agent writes 75% of its merged PRs

Come for the beautiful artsy Linear-produced video, stay for the insights about why Ramp decided to build their own coding agent. We get asked all the time why you would (or wouldn't) build your own harness. And it's certainly not for every company, but for Ramp it seemed to come down to fitting the tool to their workflows instead of the other way around:

It’s a really tight integration with our development lifecycle and our tooling

Another feature we see repeatedly from companies building their own harnesses is mentioned almost in passing in the text: shareable sessions. I'm sure we'll get there eventually with the Claude Codes and Codexes of the world, but every team we talk to on the edge seems to have shareable sessions, which fits nicely into our "multiplayer is the future" belief.

How Vercel's agents build on-brand pages

Tired of AI design slop? Vercel was, so they decided to do something about it. Now they have a public design.md file that anyone can point any agent to, turning generic AI designs into on-brand artifacts that "feel" like Vercel. Our favorite detail is that they found agents were inventing their own tokens and CSS classes, so they also published a CSS file and instructed the agents to use that, which is a chef's-kiss blend of deterministic and non-deterministic solutions.

And here's what we're talking about this week:

Seven Grok Bots Claire actually uses

We keep talking about Grok Bot because we keep finding more and more use cases for Grok Bot. This week on How I AI Claire admitted to murdering her beloved OpenClaw agents in favor of the cloud-based Grok Bots. Claire details some bots she created to help run her personal life, as well as some bots we created to keep things chugging along at CXO.dev.

GPT-6 Astra was released, and Claire is in love

OpenAI's latest model Astra was released yesterday, and suffice it to say that Claire is a BIG FAN. She's one-shotting 3D games, building Minority Report interfaces, hacking unhackable devices, and more.

One thing to try this week

Make yourself a design.md and put it somewhere everyone in your company can use. Bonus points for putting it somewhere public (yes, someone will object, but if Vercel can do it you can probably do it too). And, of course, when we say "make yourself a design.md" we actually mean point your agent of choice at the Vercel article and have it make you a version you can try yourself.

Have a great long weekend.

— Claire + Zach