You aren’t at Shopify, and you can’t tag River in our Slack. But the point is not the agent on the surface. It’s the substrate underneath, and the core bet that produced it: the session is the thing that must survive.

If you’re building agent infrastructure, here are three priorities:

Decouple the brain from the hands. The harness does not live in the sandbox. When you commit to that boundary, safety, replaceability, and observability stop being trade-offs. You will not be able to retrofit it. Make the agent multiplayer by construction. A private agent has a ceiling: the person at the keyboard. A public agent teaches every session that comes after it. The corpus is the compounding asset; the privacy of an agent thread is a disadvantage. Treat the next agent as a profile, not a platform. The cost of a new agent product should be a new bundle on the same substrate. If your second agent forces a second platform, you haven’t built the substrate yet.

The session.

In the long run, Gordon believes that places around the US should be able to pass what is known as the popsicle test, where a kid should be able to safely walk to a store, buy a popsicle, and return home before it melts. In essence, every city should be designed to be friendly and traversable to the most vulnerable in our communities. “If you start thinking along those lines, then I think you start thinking along the lines of what infrastructure is needed to make that possible, where I would feel comfortable letting my kid do that,” Gordon said. “Halloween is like a giant version of the popsicle test because it’s not just your kid, it’s every kid in the neighborhood.”

Anything less is unexceptionable.

There is a better way, and Chattanooga has long since shown the nation how. For better orworse, the internet drives much of our daily lives and the economy, and what makes the internet possible? Broadband, which is powered by those stiff plastic-sheathed cables filled with strands of thin glass or braids of copper wire, beaming pulses of light or electricity into your modem. And yet, the US has outsourced the provision of this precious public good almost entirely to de facto monopolies that keep our internet slower, more expensive, and above all, more infuriating than it needs to be.

There is a better way.

Claude (Sonnet) built a corpus of 4,176 review comments and summaries by twelve reviewers, on over 1,700 PRs going back about 20 months. A Fable agent per reviewer pulled out the patterns in what they flagged, with examples: ~300 of them, which another agent compiled into 67 themes. Then I had it turn the themes into checks. Finally it went through every PR merged since the start of September (~500) to work out which checks the version the first reviewer saw would have failed, and whether review caught it. As on many teams, there were already AI review bots on the repo, so “review” in everything that follows means humans and bots together, unless noted otherwise.

That seems worthy of a review bot

As fantastical as Disney World’s imagineered system may seem — and it was, to be clear, predicated on a not-easily-reproducible master plan — the idea of storing utility lines in easy-to-access tunnels is not just for amusement parks. It is possible even in places where most of those lines are currently lying in the dirt, as Prague proved back in the 1970s, when the city decided it was fed up with destroying its precious cobblestone streets every time a buried pipe or cable needed fixing. The city now boasts a sprawling utility tunnel network, occasionally patrolled, evidently, by a couple of robot dogs tasked with mapping out every cable.

Amazing forethought.

](https://apple.news/AlWwvQWbRS8auvbbfAJRY1w?highlight=Somebody%20has%20to%20build%20these%20businesses,%20run%20them%20and%20outcompete%20the%20ones%20already%20there.%20It%20might%20as%20well%20be%20you.)

Somebody has to build these businesses, run them and outcompete the ones already there. It might as well be you.

None of this means nobody ever reviews code. There are absolutely changes where I want another experienced human looking. An example would be a fundamental architectural change. Assuming we did a design session as a wider team, we might want to review the code as a team or agree it was implemented right, or discuss if we want to change anything. Other examples could be something crossing a sensitive security boundary, a change with a huge blast radius, an unfamiliar part of a critical system or simply something where the team says, “I’m not confident about this.”

Review By Exception, not by norm.

To be sure, there will also be a mountain of slop. Of buggy, useless, terrible, offensive software that conforms to no programmer’s sense of aesthetic or purpose and style. But there will also be well engineered, working software that solves real problems, yet hasn’t been reviewed in fine detail by human eyes.

Pretty spot on top to bottom.

It’s just a normal Monday morning. You make yourself a nice coffee, open your computer and find yourself with 7 PRs to review. You open the first one: +24506 -3938 lines, accompanied by some AI-generated description of what they’re supposed to do. Somehow, your team has made more changes since Friday than they used to make while you were away for a few weeks.

Sounds about right.

Today, even that loose connection to utility has been broken. I have enormous respect for Qwen, but I very much doubt that a 21GB quantized version of their latest model is more powerful or useful than Anthropic’s latest proprietary release.

If the thing you need is an SVG illustration of a pelican riding a bicycle though, right now Qwen3.6-35B-A3B running on a laptop is a better bet than Opus 4.7!

That is pretty impressive in 18 months.

I was telling my son’s friend about this phenomenon and he responded with the perfect GenAlpha summation:

AI is a D1 glazer, bro.

We need to acknowledge that we’re probably getting the answer we want rather than a cold-hard fact. Not to get too serious but when I read about AI psychosis, I think the overly-confident “You’re a genius” style of reply is the point where it all starts to go wrong.

Hahahaha. A “D1 glazer” indeed.

There’s an uncomfortable parallel between using AI coding tools and playing slot machines28. You send a prompt, wait, and either get something great or something useless. I found myself up late at night wanting to do “just one more prompt,” constantly trying AI just to see what would happen even when I knew it probably wouldn’t work. The sunk cost fallacy kicked in too: I’d keep at it even in tasks it was clearly ill-suited for, telling myself “maybe if I phrase it differently this time.”

For real.

Humans interact with the system almost entirely through prompts: an engineer describes a task, runs the agent, and allows it to open a pull request. To drive a PR to completion, we instruct Codex to review its own changes locally, request additional specific agent reviews both locally and in the cloud, respond to any human or agent given feedback, and iterate in a loop until all agent reviewers are satisfied (effectively this is a Ralph Wiggum Loop⁠(opens in a new window)). Codex uses our standard development tools directly (gh, local scripts, and repository-embedded skills) to gather context without humans copying and pasting into the CLI. Humans may review pull requests, but aren’t required to. Over time, we’ve pushed almost all review effort towards being handled agent-to-agent.

Didn’t know they had coined it “harness engineering”

Eliason hopes their advocacy will have helped transform American cities. “I guess success would be there’s no housing shortage, and we’ve developed super walkable neighborhoods with abundant housing and multigenerational housing,” he said. “We need this broad array of housing types — social housing and market-rate housing and condos and co-ops. Single-stair, at least, helps to unlock some more opportunities.”

Yes!!!!

He pointed out that as average humans, we had never lived the scientific process the way we did with the Covid crisis and the emergence of the vaccine. No wonder everyone from pot-smoking podcasters to scientists had differing opinions. The whole thing was playing out in front of us in real time.

The trials, the debates, the reversals, the uncertainty, the competing claims, the people who were sure they knew what was coming and turned out to be wrong. We watched science happen. It was messy, public, unfinished. And everyone became an armchair epidemiologist overnight.

Living the very messy process right now.

In a news release late Tuesday, Moderna said it was blindsided by the FDA’s refusal, which the FDA cited as being due to the design of the company’s Phase 3 trial for its mRNA flu vaccine, dubbed mRNA-1010. Specifically, the FDA’s rejection was over the comparator vaccine Moderna used.

Fucking ridiculous

Developers can leverage coding agents, including Anthropic’s Claude Agent and OpenAI’s Codex, directly in Xcode to tackle complex tasks autonomously, helping them develop apps faster than ever

🥱

The ultimate result is the ability to create software at scale without directly writing any code, simply by providing strategic direction to a fleet of coding bots. Call it “codeless” software.

This is the next real wave.

mRNA vaccines against high-risk skin cancers appeared to reduce the risk of cancer recurrence and death by nearly 50 percent over five years when compared with standard treatment alone.

Fuck yeah science!

What Made This Possible

Building AI features in a browser is not the same as building AI features in a standalone app or a cloud service. The constraints are different. You have limited resources, strict privacy requirements, and the need to work across Windows, macOS, and Linux. Here is what made it work:

Starting small: We did not try to build everything at once. The first runtime was minimal. The first model was simple. We added complexity only when we needed it.

Privacy as a requirement, not a feature: Every decision started with “can this run locally?” If the answer was no, we either changed the approach or did not build it.

Reusable infrastructure: We built the runtime once and used it for multiple features. That meant each new AI capability got cheaper to ship.

Learning from real users: The validation app for PDF.js alt text was not just about improving the model—it was about understanding what real people needed. User feedback drove every iteration.

The return on investment has been extraordinary: the network has delivered 6.4 times the value of its original $396 million investment, the study indicates.

💥

Their code pretty much reflects this philosophy: Plain, boring and concise with no fluff. You can easily read it top to bottom without needing to jump around to get the bigger picture. There is very little noise there. There is this pragmatic minimalism to it. You won’t find layers of abstractions for abstraction’s sake. The biggest “debate” - if you could even call it that - usually revolves around how to structure data or where to put side effects.

Clarity is key indeed.

The nonprofit RSL Collective brings together millions of online publishers and creators to protect their rights and negotiate for fair compensation from AI companies.

Definitely something to keep an eye on moving forward.

Zoning rules prohibit the construction of apartment buildings on roughly 75 percent of America’s residential land. Throughout much of this territory, land-use laws effectively require all single-family homes to be spacious (and thus, pricey to build or buy). Even in city centers, parking mandates often make multifamily housing financially or physically nonviable.

Just nuts. No great mystery why there isn’t enough housing. 🙄

The app has had a major overhaul to all areas while keeping a familiar UI. Base 3 is faster, more capable and even nicer to use.

Very nice.