Hamilton vs the race director. A tale as old as time

A few of us traded keyboards for stadium seats Monday night—Yankees vs. Tigers. ⚾ Team culture goes beyond surveys and offsites. Hanging out outside of work makes building AI infrastructure together that much more fun. Don't forget, you can't spell "happiness" without "API." @AlexTriplett @patrickdonohoe @WallensteinTom @seherdholakia

Pioneering the web for the agentic era

First-class integration!

Reliability 💪

Making our APIs more accessible!

For most industries, AI won’t transform work by doing every task directly. The bigger shift is AI making it cheap to create bespoke software that automates work and evolves over time. Most actions should be deterministic. AI is the layer that builds, maintains, and improves the system.

I don't read the code for my PRs. i ask my agent to explain all of the changes to me. and then i provide feedback. this is going to be how everyone develops software in 6 months

everyone should have one claude code session that is your team lead and responsible for coming up with tasks to give to other ic cc sessions

using fable again is like hiking to the top of a mountain and breathing out of an oxygen can. euphoria

lets go!

in the ai era, solving a single challenging part of the stack lets you go into adjacent layers in ways that weren't possible before. look at recent news with waymo and uber: waymo is cutting uber out of rideshare and going direct to consumers

it almost pains me to work with my current coding agent loops. I see so many places where I am just burning tokens. but based on the current limitations of the harness the leakage is required. i see us getting more customizable harnesses and improved ways to feed our agents memory context so that we can dedicate more tokens to shipping

anthropics loop agent got leaked. I think it reinforced one thing that I have seen building my own loops. having a robust verification step is the most important part. this is why you need to lean in heavily to to config-as-code and away from provisioning infra with a cli. clis let an agent go rogue. config-as-code gives you a verification step where either an agent or human can review work before its merged to prod

hyperscalers are investing all their free-cash-flow into ai right now. its a moment in time to lean in on the company and personal level

one thing i have noticed about cc vs codex. recently cc has leaned hard into smaller diffs for PRs. It does lots of slices for a feature. I bet this is because the smaller slices with clean tests and checkpoints are better for accuracy. I have started to teach my codex to do the same thing

it would be great if there was a keyboard shortcut to change the effort level in the cc desktop app @bcherny

there has never been a stronger signal to go for a morning workout

one thing that i find very interesting about the open-versus-closed source debate. up to this point, open-source has largely been funded by the chinese gov't to push them to the frontier of ai. and it has been very successful. But now these companies are raising more institutional capital. There is going to be pressure to make a business out of their incredible research labs. We already saw what this means at meta. Will these companies be able to give their models away for free in perpetuity or will they be pressured to go closed-source? Could have an impact on inference providers