Being a dev is interesting.
In one moment, I'll be looking at work another dev is doing and say "Wow, this stuff is so cool ... I feel like I need to step-up my game"
... and the next moment I'll look at work (often by the same dev!) and think "Woah ... um, I'm doin' just fine"
Started streaming Blade Runner on Amazon Prime. Strangely hear Harrison Ford giving voice-over narration to his thoughts ... wtf?
Apparently, that's the Theatrical Release version. I've only ever seen the Director's Cut and Final Cut.
Wow, that voice-over completely ruins the atmosphere.
So why'd they add it? Because when screening the movie for audiences before release, mid-wits were "confused". This is why we can't have nice things!
It reminds me of Pierce and Inspector Space-Time (IYKYK).
In 2023, I spent hours upon hours hand-writing code for one major enterprise.
In 2026, I spend hours optimizing teams (of AI agents) as THEY write code for multiple profitable businesses.
The crazy thing about it is ... every step I have taken just felt natural. I have no sense of "omg, the entire industry ... and the entire world ... has changed completely". It's just been "Hey, new tool. Let's see how to use it."
Weird how that works.
Eating meat = ♥️😀😀😀♥️
Was sluggish for a few weeks. Five days ago, decided to start eating A LOT more meat (esp. red meat).
Feeling GREAT now. Also, MUCH happier.
I don't just need daily protein, I need 5 daily steaks.
I've just published a first version of Heimdall, my gate/QA repo, to GitHub. npm package is under review and should be released shortly as well.
What does it do? Why do you care?
It's a repo I wrote for my own workflow. It passes a package of common QA jobs (ESLint, type-checks, duplicate-code checks, dead-code checks) to a script that runs and returns a summarized result.
1 - It helps strongly discourage AI agents from 'cheating' their way through these checks or skipping any of them, as it's a deterministic script.
2 - It saves tokens (for me, a lot of tokens) by offloading these jobs from AI to a script and returning a summary result. Now AI doesn't oversee and parse ESLint checks or knip runs ... and it doesn't waste the tokens or time doing so (yea, the script runs faster than it would through AI as well).
3 - Running these checks helps keep code quality up.
It may or may not be the right solution for you ... we all have our ways of doing things ... but it works quite well for me and I promised to share it in replies to one of my posts on Saturday, so here it is:
(At mentioning people who directly expressed interest: @AlxTurovski@JSchiezty@proximalcoast@Wv3EcACHtizRjZKgithub.com/kahlstrom/heimdal…
Years ago, resorts and restaurants got so tired of "influencers" demanding free meals or free stays in return for good-reviews that a whole backlash against them became part of the zeitgeist.
I was happy. Influencers have always annoyed the hell out of me. Sadly, that backlash was extremely short-lived. Now, entire companies exist for the purpose of paying influencers for good reviews. We're stuck with them.
I searched my emotions about them to figure out why I hate influencers so much, ideally to remove my own prejudice and 'just get along' ... but I can't. Influencers are fundamentally built on:
> "Building a brand", "gaining popularity" instead of authentic presentation of self
> "Prestige speech" instead of communication for the purpose of exchanging real information or discovery
> Dishonestly selling ones-self out to promote in return for $
All three of the above things are antithetical to the 90's punk rock / skater / snowboarder culture I grew up in. Maybe not to all various brands of this culture throughout the US or the world ... but in my particular area of the Rocky Mountains, all three of the above items would get you shamed, expelled from the group, and possibly even physically assaulted. And I still believe that way. With my entire soul.
Authenticity matters. Communicating honestly without 'prestige speech' matters. Moral-signaling through buzzwords is just a sign of vapid stupidity. Selling-out is never right ... ever. Selling-out is the definition of abandoning your soul for money.
What's the point of my rant?
I'm not really sure. I was just engaging in thought exercises of what might replace influencers in the future and ideating on "how can we get there faster?" ... and all of this tumbled out of my brain here onto X ... which is how i do all my posting. It's just what I'm thinking about in the moment.
How I handle QA with my agents:
lvl 0: Story-building with my Grok Bot PMs that takes a feature I need, breaks it into discreet, easily-held-in-context, stories, and then feeds those stories one-at-a-time to my coding agents through an entire feature, without me getting involved.
lvl 1: Global Skills and Rules for my coding agents in my IDE, as well as repo-specific AGENTS.md that handles anything that needs to be only on that repo.
lvl 2: A 'gate' small program in a repo of its own that is called by every AGENTS.md on every repo and runs unit tests, eslint, dead code checks (with knip), and duplicate code checks (with jscpd). It does this after every story is coded. Failures are immediately sent back to the coding agent to fix. If there are more than 3 rounds of fixes, I get notified to see what's up.
lvl 3: A QA Grok Bot that does final checks, architecture review, etc after the entire feature has been written (all stories passed through gate checks and are completed). Failures result in pass-back to the coding agents and after 3 failures, I get notified to see what's up.
This process is the result of several weeks of working with Grok Bot as the layer I interact with and keeping myself out of direct interaction at the coding agents layer.
The 'gate' program was built to help reduce token usage and keep a deterministic testing/gate process for each story that doesn't vary or fail to check something. If I shared it, would anyone be interested?
I'm at 80% of my weekly usage allowance -- on a Cursor Ultra ($200/month) plan. I have 4 days remaining until reset.
Granted, I've run coding for ~12-14 hrs/day every day for the last 3 days ... on 3 projects simultaneously.
Time to see if I can make things more efficient.
Today we’re introducing UI SDK.
Agents are becoming both the builders and the users of interfaces. So we asked ourselves: where should the abstraction live?
Take a Select. There’s a popup, selection logic, keyboard navigation and focus management. Those same behaviors are useful in lots of other components.
So we pulled those pieces apart, tested them, and made them reusable. The behavior stays the same, whichever framework, renderer you build with or the harness generating the UI.
That’s the idea behind UI SDK. Give agents the pieces to generate the interface you need, right when you need it. Even interfaces we haven’t thought of yet.
I've setup my agents in a way that lets me be so hands off that I'm starting to feel guilty about not doing enough work by hand myself.
And suddenly I understand why so many middle managers spend all of their time calling meetings, asking for updates, and being annoying.
🤦♂️ My agents were having a hard time getting large asks right over the past few days. They could handle small asks just fine ... but bigger ones would come back with short-cuts taken and flags for missed implementations. I was wondering WTF was wrong.
I was using Composer 2.5.
Don't get me wrong, Composer 2.5 is a great model ... but it's never been good at large tasks and has always needed the user to break things down into small, manageable, pieces.
Whoops.
An update from Elon on Grok 4.8. Things I like about it:
> They recognized that Grok 4.7 was crap
> They are doing something about it
> We're going to see Grok 4.8 this week on Grok Bot (and I imagine that means next week on Cursor)
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.