Years ago, my wife's grandmother once asked me what I do for a living. I told her I build websites. She thought about it for a second and then asked, "Like at a factory?" Everyone laughed. Turns out maybe she wasn't wrong, just a little early.

I was reading The Art Thief recently so I have hand-crafted art on the mind. Before industrialization, if you wanted a table, someone made you a table. They were crafted by hand and were expensive. They also were built to last.

Then manufacturing came along and suddenly furniture was cheap, accessible, and functional. You can still tell the difference between an IKEA bookcase and something a craftsman built, but for most purposes IKEA gets the job done. Code has been on a similar trajectory for a while. Component libraries, design systems, npm packages...we've been reducing the per-unit cost of code for years. Agentic coding is the step that makes "factory" feel literal. Compiler throws an error, the agent reads it, tries again. You go get that third cup of coffee. You come back to a working thing.

Lowering the cost of production doesn't automatically raise the quality of what gets produced. The engineers who stand out are going to be the ones who develop taste. Real taste for what makes a good product, a good user journey, a good design. And increasingly, understanding what agents do well and where they fall apart, and engineering the processes around those realities to get the best outcomes. The model isn't the hard part anymore. It's the taste and the craft that goes into building systems that produces good code.

Let's dig into this week's links.

💻 Tool
Code Mode
Cloudflare
Cloudflare's Code Mode collapses its entire Workers API — thousands of endpoints — into two tools: search() and execute(). Instead of loading the full spec into an agent's context upfront, the agent discovers what it needs as it goes and runs the processes on Cloudflare's Edge to give the model the context it needs.

🤯 Tool
Markdown for Agents
Cloudflare
Staying with Cloudflare they also recently announced Markdown for Agents. Their feature that allows agents to request text/markdown versions of pages, dynamically rendered on Cloudflare's network. This can take the token cost of a blog post from 16,000 tokens to 3,150. It is a huge reduction and gives agents the ability to do more before having to manage context.

📰 Article
Building an Agentic Memory System for GitHub Copilot
GitHub
This one slid under my radar — and we use Copilot at work. GitHub built a cross-agent memory system where agents store what they learn about a codebase, with citations back to the exact file locations. Future agents verify that knowledge is still valid before applying it. It's a thoughtful answer to one of the more persistent frustrations with agents: they forget everything, every session. The numbers are modest so far (7% improvement in PR merge rates) but the concept is sound. The model's the same. The memory layer is what's new.

🐶 Article
I Taught My Dog to Vibe Code Games
Caleb Leak wired his 9-pound cavapoo Momo with a keyboard, routed the random keystrokes through a Raspberry Pi, and had Claude Code interpret them as "cryptic commands from an eccentric game designer." The result: actual playable games. The dog's input didn't improve. The system around the dog did. Caleb gave the agent the ability to screenshot its own output, playtest its own levels, and lint its own scene files. Not to mention decode the gibberish input into requests. A dog made working video games, which is mostly a fun story, but also the clearest illustration of the point in this entire issue:

The expensive part isn't writing the code. It's knowing what the agent needs to succeed and building those conditions intentionally. There was real work done to set that workflow up. None of it happened by accident.

🔧 Tool
Wispr Flow
I started playing with Wispr Flow last week to try out its voice to text model. And it has been surprisingly nice. I've found myself triggering it to send off a prompt rather than type one out. It may just be I've had one too many undecipherable Siri reminders made during my morning commute, but Wispr's auto-editing and capturing makes for a better brain-dump experience to get info out of my head and into my notes. Eventually we’ll all just be coding with our brain chips anyway.

Thanks for reading,
Jason