Thomas Sievering

@badlogicgames Will come if you give me more fatherhood / parenting advices!

who here will be a AI Engineer London in April? I'm ready to have more pub visits.

@badlogicgames A way to find out more about said workshops?

i give agentic coding workshops every now and then. i have a demo where i let the clanker clone the amp landing page in various ways. just found out their landing page changed and i have 20 minutes before the workshop starts :D

we're here already

@grashalm_ @johncrickett yeah, the tool probably doesn't travel. the habit of making the tool does.

@sieveringthomas @johncrickett In the past we just didn't think about one off tools for many things. I do not want to share some of them, because they are too personalized for anyone else to use. Specialized code is often much less complicated. I think most personal tools won't get reused in the future.

12 months ago I said this wouldn't age well. So let's check in. AI writing 90% of code? We're not there. Not even close. Sure, there are a few bubbles where adoption is high, in the AI companies themselves and across Silicon Valley. The rest of the world, not yet. Here's what I actually see across the industry: → Some developers don't use AI at all. → Many use autocomplete, and that's it. → A good chunk use chat interfaces as a smarter search engine. → Around 20% are using agents, and mostly just for writing tests. → Around 5% are heavily using AI. Personally I use it a lot. It's genuinely useful for the work I do. I'm in the 5%. But "essentially all the code" by now? That's not the reality most software engineers are living in. The gap between what AI leaders predict and what practitioners experience on the ground is still vast. That doesn’t mean it’s not useful. That doesn’t mean it’s not here to stay. It's just a reminder that timelines from the top of the hype curve tend to compress years of messy, uneven adoption into a single clean narrative. Or as William Gibson put it so well: “The future is already here – it's just not evenly distributed.” Which group are you in?

@MingtaKaivo write a style guide for the agent. 'match existing patterns, no new abstractions' is uncomfortable to write but once it's there, taste becomes spec.

@sieveringthomas solid point. acceptance criteria works for the mechanical stuff. but the CORS fix wasn't hard—my agent just couldn't decide if the solution was clean enough. that's the gap. knowing 'it works, ship it' vs 'let me try one more approach' is still a human call

agents are terrible at knowing when to stop iterating. watched mine run 47 refinement loops on a CORS fix before I killed it. humans know when the 80/20 point is. that instinct isn't learnable—it's taste

@coolnalu types are machine-readable specs. the model doesn't infer worse, it just has less to work from.

I feel Claude Code is much more clumsy on dynamic typed languages - i.e. python.

@imfrancoierace skill files are definitions, not constraints. limiting what the model does is different from showing it what good looks like.

your job now is to automate your job

@bcardarella the gap isn't tooling. it's whether you defined what 'done' looks like before the agent ran.

To be clear, I think there is a big difference between "vibe coding" and "agentic coding". Coding with AI is legit but you cannot one-shot your way to success

@grashalm_ @johncrickett fair. but the 'built for yourself' label doesn't last. most personal tools end up in someone else's workflow before the first PR.

@sieveringthomas @johncrickett It depends. If u are building long term maintained software, then definitely having a full review is mandatory. If it's a tool you built for yourself then it's fine to not even look at the code most of the time.

@skirano the test is simple: does it end in a git commit? if it doesn't, all you've done is move the translation problem upstream.

Doubling down on this: the recent wave of AI design tools has done more harm than good. None of them work with real code. Using AI to create yet another layer of abstraction is exactly the wrong move. Design is code, and code is design. And you should care about interactions.

In a few days, we’re shipping something at @MagicPathAI that makes the dream of designing in code with perfect handoff, especially for large teams, finally real. I’ve been in this field for 20 years, and early in my career this would have felt like an ontological shock.

@plainionist true for the whole loop. what to build, how to verify, when to stop - still on you. leverage is in how precisely you spec those boundaries.

You can let AI write your tests. But you still have to engineer them. You decide: - what to test - how to test - the QA strategy AI generates the code. You engineer the quality.

@grashalm_ @johncrickett diff format as the review gate is underrated. you're forced to see what changed, not just trust the summary.

@johncrickett Agree. But tbh swarming agents only works for a small subset of my tasks. I do let the agent push the first version directly as PR. And then I iterate with it there. Let it respond with a commit as change. It is simple, persistent and when needed collaborative.

@fhinkel the tell is whether anything runs when you close your laptop. most people claiming group 2 are still babysitting prompts.

There are two types of developers emerging: 1) Those using AI to work faster. 2) Those using AI to build systems that work for them. The first group is vibe coding their way to obsolescence. The second is loop engineering the future. The choice is yours. If you are just a faster version of your 2023 self, you are still a manual operator. True leverage comes from building autonomous loops that scale without your intervention.

@headinthebox the eye roll usually comes right before the incident. then suddenly everyone wants your number :)

@sieveringthomas Absolutely. But whenever I pitch this I get eye rolls and are treated like I am a dimwit.

"... But the model literally caught the attack and described it perfectly, after the payload had already executed! ..." Perfect real world example of why we should do safety checks on a representation of a plan *before* anything gets executed. cacm.acm.org/practice/guard… Who is rolling his eyes now ;-)

@mitsuhiko cost to ship dropped. cost to launch dropped even more. the ratio of signal to noise got worse, not better.

I'm trying to catch up with some recent agentic engineering developments and I keep being reminded by this tweet. Barely anything these days is just people sharing ideas, people share products and vibeslopped websites. So much effort goes into marketing. x.com/mitsuhiko/stat…

@solarise_webdev @GeoffreyHuntley yeah. the ones promising UBI are usually the same ones who'll be fine either way :)

@sieveringthomas @GeoffreyHuntley The constant promises of “UBI utopia” are naive and patronising at best Feels like very few people with the wealth and resources to do so are trying to solve real problems instead of vanity projects or trying to cement their place amongst the stars.

it’s undeniable that things are changing at break neck speed but the question remains where is this all heading? anyone who says they know for sure, is full of shit, at this point all I can kinda do is provide colour to binary thinking and pattern match.

“porting software has been trivial for a while now. here’s how you do it” (link below)