Ranger ads

What do you do when a teammate ships 20 PRs in a day? That came up at a dinner Ranger hosted in Denver last week. Not as a hypothetical, but as an actual problem people in the room were dealing with. Al Kemner, David Guckian, Anuj Biyani and Gordon Dri joined me and Daniel Griffin for a few hours of conversation. We covered a lot of ground: foundation model training vs. post-training, where LLMs still struggle outside of code, and the AI hiring/interview questions nobody has fully solved. The best conversations I have are the ones where someone says "honestly, I don't know." We had a lot of those, and we're excited for more!

The Pope is now on Anthropic's side. Not the hire I saw coming but honestly the resume is there. Future is blessed!

The name "Ranger" was always about more than testing. Rangers navigate. They know the terrain so you don't have to figure it out yourself. They make sure you can go further, faster. That's what we've been building, and now, our new homepage shows it. There's a detail worth finding: the background image is the national park closest to wherever you're reading this, because we built a geolocation feature that surfaces it automatically. Let us know what you think!

I gave a flash talk at a bug-bashing conference hosted by Antithesis a few weeks ago. The one thing I tried to land was that reliable software and good software are not the same thing. The correctness community is solving reliability. Fuzzing, formal verification, deterministic simulation — does this code do what it's specified to do, every time? That's a real and important problem. But that isn't the same as: Does the software do what users actually want? Does it make the decisions you would make? Does it feel right? As coding agents generate more software faster, this distinction is going to matter more. Your factory doesn't just plateau at generation speed. It plateaus at the speed you can review for goodness AND correctness.

I spent yesterday at Code w/ Claude (and Boris Cherny) in SF. The throughline across almost every talk I went to was that the model is not the product. The harness around it is. Dario Amodei's keynote put a name to it: Amdahl's law hitting AI. As generation speeds up, the bottleneck just shifts to the next slow thing. Right now that's verification. Jarred Sumner from Bun made that point more concrete than anyone, and his team has been living this problem. He started with a simple exercise. He asked everyone to raise their hand if they use Claude Code in the terminal. About half went up. Then he asked who has a cloud agent running 24/7, automatically reproducing bugs, writing fixes, and generating proof that the fix works. Almost no hands. That gap is where the next few years happen. To show what closing it looks like, he walked through something they built called Robobun. It pulls issues from GitHub, reproduces them, writes a fix, then generates tests that fail on main but pass on the fixed branch. The test isn't just the artifact, it's also the proof. His framing for all of it was: "The new bottleneck is, do I feel good about merging this?" I've been thinking about it since because if you write a test you know needs to pass, an agent can hill climb until it actually does. That's clean for unit tests. What that looks like for end-to-end flows is the next problem... Generation is mostly solved but what comes after is not.

"Making the tests pass" and "making the code work" are increasingly two different jobs. AI agents (yes, even Claude Mythos) are speedrunning the distinction.