Luke

near-determinism

want to start talking about agent workflows in the context of specific stacks, rather than just generally “this is how to work with agents”

🗞️ nobody should type code by hand anymore. here's what's left. ghuntley.com/agency/

my workflow for the last few months has been: spend 2-3 weeks thinking about/designing something and then build it in a few days with minimal iteration. then it’s immediately onto thinking about the next big/interesting problem for 2-3 weeks and build it in 2-3 days

Code that works at a point in time is mostly uninteresting. What matters is building structures to ensure code stays correct over time. That to me is the essence of engineering in software

If you're curious how all those newfangled domain names come to be -- I wrote a feature for WIRED on top-level domains and the people who make them I thoroughly enjoyed meeting all the kind and fascinating people in this wacky domain world. (And I promise it has nothing to do with cat thongs.)

why did we think it was a good idea to pay money for an agent to call an API? we’ve been calling APIs for decades, for free!

like this if you have realised software is better than chat

100%

Claude Code explaining why it changed three files.

one shot culture is weird www.quarter--mile.com/One-Shotted

still trying to answer this question in 2026 (50 years later)! "there is a critical question facing software engineering today: How to modularize a software system so the resulting modules are both testable and maintainable?" A Complexity Measure, McCabe - 1976 literateprogramming.com/mccabe.pdf

Claude Opus 5.5 is out. “It's ~30% faster and ~40% cheaper than Opus 5 per task. In Claude Code: - 5-hour session limits increase 20% today - Opus 5.5 is priced lower, so it goes 25% further within limits - Pro, Max, and Team users get a reset to use anytime” www.anthropic.com/claude-opus-...

whatever happened to agentic commerce?

this is what my team has spent the last few months doing. the one thing in software engineering you never used to do is now the best thing you can be doing! the results are incredible - you can finally have the software you need to support all the features you want to ship!

fascinating series tbh. And honestly, I too would like to drink coffee and work on cool tech in a comfy mind upload palace while fighting aliens and traversing wormholes….

ancient puck wisdom: “the ceremony is defensible, but it does insist on being fed”

agents orchestrating agents? you’re doing it wrong durable, deterministic and efficient agentic engineering needs software calling agents

Funghazi

bsky.app/profile/luke...

As a coder, you need at least 2 AI providers in case one is offline. Use different models from those providers for different tasks.

slide from a deep mind talk at the Barbican in London circa June 2019

every damn time

most people probably don't understand that agentic software development is now effectively the industry standard, even to the point that Zohran Mamdani's administration leverages agentic software development this is the reality we're working with, and it's a very tricky environment to navigate

TL;DR is basically that I don't want to use LLMs to gain speed (e.g. add more features faster), but to increase robustness and quality (and I would even accept *losing* implementation speed for that).

Frontier AI models are converging on similar intelligence, yet cost per task can still vary 10x-60x across vendors. That seems unsustainable as models become more substitutable. A lot of frontier lab economics depend on customers paying huge premiums for increasingly marginal intelligence gains.

can’t decide if this is nowhere and it’s now or land locked blues

does the same existential threat apply to internal platform teams too? if product teams are best placed to design tools to fit their use and can now build them to exact specifications, the generic solutions platform teams provide are no longer worth the hassle

I am not pro AI lsh.pckt.blog/i-am-not-pro...

Multiple people I know have told me that they love the idea of Bluesky, and want it to succeed, but have abandoned it for X because the use agentic tools in their work and find them incredibly useful, and feel that any mention of their usage here leads to hate and ridicule.

AI is always going to be polarising but it is a shame that we can’t have considered debates yet - it always has to be “it’s perfect” or “it’s evil”. Neither extreme is true and it’d be nice to chat about the spectrum in between on here

Nope! But swamp is growing a community super fast :)

Load-bearing seams relitigate the honest takes, and that’s where the ratchet bites

have you seen anything else that looks like swamp emerge yet @adamhjk.me ?

wrote up some thoughts on the yegge Thunderdome post build your own harness, CD is dead, code review is long dead! lsh.pckt.blog/harnesses-ga...

What’s one side project you moved to serverless/Workers that you’ll never go back to traditional hosting for? #CloudflareChat

ooo this looks like a good rival to the GitHub Copliot app, which I’ve been using lately as the main driver of agents. the multiplayer aspect will be super cool to see how it feels let me in @zed.dev !

the existential threat to companies building ai-based devtooling seems to be personal software devtooling appears to be heading towards bespoke software written by you/your org for you/your org build vs buy for devtools will settle to build if you need and can design it, your agent can build it

if you haven’t read the latest from yegge you should. “I don't care how much money you have; your spouse is not going to let you spend almost ninety thousand dollars a month on your hobby game. I didn't even bother asking.” yegge.ai/essays/the-s...

@cloudflare.social is blog.cloudflare.com/cloudflare-o...

this is great advice!

do you really need a platform team to manage IaC…?

who is working on deterministic automation for agents?

wonder what tj holowaychuk is up to these days