Phil Calcado

Reviewing a 4,000-line agent-generated PR in 2026 be like:

Jev and the System One models in its wake: you either get it immediately, or you can’t see why anyone’s excited about a model that can’t write a sentence.

If you get it, you’ve probably spent years torturing LLMs into production.

Slides from today’s session: https://philcalcado.com/downloads/jev-and-system-one-models.pdf

Twitter is Skype
Mastodon is IRC
Threads is MSN Messenger
Bluesky is Yammer
I will not elaborate.

AI commoditizes SaaS the way VB and Delphi commoditized 90s CRUD apps. When building gets cheap, design is the differentiator. Yet most of what ships is commodity agents writing commodity code behind commodity UX.

Katie Dill’s is a breath of fresh air https://www.youtube.com/watch?v=GLvFTMtw4Jk

We might have used up all the world’s water, but we finally got DHH to stop writing code.

Worth it.

For over ten years I've run the same structured RFC process for how engineers propose and debate technical decisions. In just a few months, AI broke a foundational assumption: that a serious document meant somebody had done serious work.

So I wrote an update for an AI world—as an RFC, obviously
https://philcalcado.com/2026/07/10/writing_ai_assisted_rfcs.html

Nobody writes a $3M seed check for a classifier, so at my last startup we called ours something else.

What was actually under the hood, and why TypeSafe AI's launch caught my eye:

https://philcalcado.com/2026/09/15/classification-is-often-all-you-need.html

Being Brazilian in an American coffee shop means everyone else gets pleasant bossa nova ambience while I’m forced to understand that this man has been singing “sun, love, sea, you” on a loop for three minutes

> we’ll need senior engineers. People who understand systems end to end, who can debug distributed failures at 2 AM, who carry institutional knowledge that exists nowhere in the codebase. Those engineers don’t exist yet because we’re not creating them. The juniors who should be learning right now are either not being hired or developing what a DoD-funded workforce study calls “AI-mediated competence.” They can prompt an AI. They can’t tell you what the AI got wrong.

https://techtrenches.dev/p/the-west-forgot-how-to-make-things?triedRedirect=true

when the coffee shop puts on a bossa nova playlist and you’re the only Brazilian within a five-mile radius

Being Brazilian in an American coffee shop means everyone else gets pleasant bossa nova ambience while I’m forced to understand that this man has been singing “sun, love, sea, you” on a loop for three minutes

Captcha-Driven Development

The hardest part of being a senior engineering leader is making enough room to do the job

https://philcalcado.com/2026/08/18/treat_escalations_as_defects.html

Email wasn’t replaced by Slack but by four hundred Google Docs on the same topic, all named 'Untitled'.

I mean, if we’re writing memos like it’s 1954, we could at least bring back the pneumatic tubes.

When Leftpad hit the spotlight, we couldn’t believe anyone imported a lib to save 11 lines of code. Now we burn a million LLM tokens just to indent a file.

shut up, claude, nobody cares

Lots of thoughts and feelings lately about the whole "code was/wasn't the hard part." I think it's a valid discussion, but only if we know what kind of code we’re talking about...

Commodity code, which is what the majority of SaaS applications are made of, is rarely the hard part indeed. Unfortunately, this often also reflects on the products being commodity...

I wrote about this distinction a while ago, and I think it’s even more relevant now:

https://philcalcado.com/2025/02/27/ai.html

If anything, the constant pivoting and re-pivoting to build an AI narrative has probably slowed down the delivery of real value while increasing noise and feature churn.

I wrote a short thing--it's been a minute

> This creates a loop of Brandolini’s law: not only is the original bullshit cheap, but even if you spend a disproportionate amount of time refuting it, generating a new bullshit response to your comments is even cheaper.

https://philcalcado.com/2026/03/06/ai_and_brandolinis_hell_loop.html

It's still hard to predict how dev tools will adapt to AI development, but one thing that's already pretty clear is that JetBrains and other tool vendors need to beef up their support for Git worktrees and other ways to coordinate multiple streams of work over the same repository.

And with GitHub degrading every day, might as well gear up for the next wave of code hosting while they're at it!

First WFO (Work From Outside) day of the year

bad ending: the whole world is scrum master

The greatest trick the devil ever pulled was rebranding consultants and sales engineers as FDEs.

This generation’s Y2K starts when subsidies end and everyone tries to save on tokens.

The management and incentive thinking I hear from top VCs today sounds like it came straight from an IBM middle manager running a mainframe payroll system in 1987.

Wait huge!
Need correct.
I almost gave wrong.
天天送钱

I appreciate what @surf@flipboard.social doing and will keep using the app, but I am disappointed that the @flipboard@threads.net team who built a mind-blowing app in 2010 wasn't able to build a better UX here. It's not just about being slow, it's that it's a lot of clicking around to find content--there's no flow

Since 2023, I’ve been bothering anyone who would listen with the same idea: AI will only really deliver on its promises once it’s in the hands of builders, not fundraisers or course sellers

It’s still early, and things could shift, but Mario Zechner’s Pi feels like validation of that idea.

🧵

Sad to see Clockwise get sent off to a Salesforce farm upstate.
Was always a big fan; it saved my ass running large engineering orgs. Ended up rolling it out across DigitalOcean and SeatGeek, and we somehow even became featured customers in their farewell post… which, in hindsight, might explain things. https://www.getclockwise.com/

Anthropic: 100% of our code is written by Claude Code

Also Anthropic: We've been running Code Review internally for months<...> 84% get findings, averaging 7.5 issues <...> less than 1% of findings are marked incorrect.

I've user PagerDuty for, what, at least 15 years?

I have no idea what any of the buttons in their terrible UI does except ack and resolve.

From what I’m seeing of Jev in real-world applications, a lot of us are about to learn the hard way that you can’t just multiply probabilities 😬