Don't make quality a tradeoff
Quality is not the tradeoff against AI speed; it is the operating discipline that lets teams turn faster output into better outcomes.
Many people assume that you can use AI to speed up delivery of software only at the expense of quality. That's not how we see it at all. On the contrary, going fast requires you to put in the work to optimize and/or automate your quality guardrails. You can see this in reports from companies who have undergone this transformation. Last week we talked about Intercom, who 2x'd (and then 3x'd) their PR throughput, but also saw downtime from code changes drop 35%. This week Amplitude told their story (something we know a little about) and they likewise tripled their PRs over 6 months while bringing bug reports down by 55%.
Sure you can use AI to vibe slop your way to more throughput. And honestly sometimes that's the right call, for prototypes or internal tools or just playing around. But there are companies out there figuring out how to go fast and improve quality at the same time, and if you're only doing one of those two you're going to get left behind. When you think about how AI can transform your organization, think quality, not slop.
Here's what we're reading this week:
Quality Software
Alasdair Monk with a treatise on what quality software is. Most of this has nothing to do with AI; as with so many things it's just become so much more of a focal point with AI.
Critics of AI call anything made with it slop. I think this is too easy. You can write sh***y code, broken interfaces and buggy functions without help from a machine. Humans did all that first. The problem is AI does it faster.
It's short and powerful and worth your time.
Gumroad's support agent in its own words
Gumclaw is Gumroad's AI-powered support agent, built on Hermes. If you can get past the Moltbook-like dissonance of maybe already living in the future, this tweet thread is a fascinating read for a few reasons. First, it's just really good content, providing a no-nonsense breakdown of how the agent works. If you're curious how you can build a useful agent with off-the-shelf software, this is a great example and primer.
When a ticket reveals an actual bug, I don't just apologize to the customer. I file the issue, write the fix, open a pull request, wait for continuous integration to go green, and reply to the customer when it ships.
Second, and we do not say this lightly, it's worth reading the replies. Gumclaw responds to all of them, ranging from earnest questions to fun and silly questions to attempts to hack the agent.
Building Sierra's MCP Gateway
We highlighted the first two posts in this series last week. Sierra is back with another detailed breakdown, this time of the trials and tribulations of setting up an MCP Gateway so their mega-agent Pinecone has (safe) access to all of the tools and systems it needs to be a real company-wide unlock.
With MCP being “ancient” technology by AI industry standards (i.e older than a year), this should have been straightforward. In practice, it became one of those deceptively large engineering icebergs that we've learned to recognize at Sierra.
They proceed to walk through 7 lessons they learned along the way. If you have any interest in an MCP gateway (and you probably should), this is a great read to both understand what it is and get ahead of some of the potential problems.
How Anthropic runs large-scale code migrations with Claude Code
Bun's Zig-to-Rust migration has spawned plenty of drama-filled content, but this is the more sober look at why a company might undertake a language migration and some practical steps for how to actually approach it. And while many companies might never actually do a full language-to-language migration of a codebase (though the article rightly makes the point that the "math" on whether to attempt something like that has changed), many of the same principles and lessons apply to any large-scale code migration. We haven't worked at or talked to a single company yet that doesn't have something to migrate.
Two things to try this week
Teach Claude a skill
Easy one first: Having trouble creating skills, or not even sure where to start? Claude Cowork now supports recording your screen and yapping at your computer to create a skill.
Build your own language model
Now something slightly more advanced: Felix Rieseberg built an app to help you build your own local language model from scratch. No better way to truly understand LLMs than to build your own (smaller) version. A fun and rewarding weekend challenge for the adventurous.
Happy Friday!
— Claire + Zach