SOTA is becoming a luxury tier. I think that is the cleanest way to read what has been happening lately. State-of-the-art models were always too expensive to serve at scale and now that bill is finally due.

Still buzzing from my time at AndroidMakers 🇫🇷 by droidCon! It was awesome to share about edge-AI, to catch up and listen to exchange ideas with people building great products! Thanks to @gregfawson.bsky.social and all the staff for making this happen! Till next time! ✌️

The 🤖 Android team just released three things that are a great example of what your team or company can build too: - Android Skills - Android CLI - Android Knowledge Base developer.android.com/tools/agents/

This week @kau.sh and I had the pleasure to chat with Mitchell Hashimoto on the pod🎙️! Refreshing to hear someone of his caliber bring such a grounded perspective to agentic coding. We're super proud of this one!

I’m at Android Makers 🇫🇷 this afternoon talking about “Practical On-Device AI on Android.” Excited to share about features unlocked about this new capability, reliability, hybrid approaches, models, and more. If you’re around, come join! I’d love to meet people working on similar problems.

I can't even @benkadel.com 💀💀💀

Claude Code’s source code leaked last night. Given the speed and sheer volume of forks and spin-offs already popping up, agentic harnesses are about to surge really hard. It’s crazy out there and I actually love to see it 😈 github.com/chatgptproje...

this one was fun to make. @iurysouza.dev and I took a crack at explaining how image models (stable diff.) work in ~20 minutes. building an intuition around AI is starting to matter a whole lot and that starts by understanding the mechanics. hope you get as much out of it as we did making it!

If you want to get some intuition behind how some of the image models out there work, this episode is for you! @kau.sh and I had a go to try and break it down in a short episode. No need for a masters in deep learning to understand some of the clever tricks behind image generation. Check it out!

If you're unsure what model your machine can handle, you can check axjns.dev llmfit. It's a TUI that profiles your hardware and scores models to tell you exactly what fits in your RAM/VRAM. This is it running on my mac:

Issue 04: Openclaw, Cowork and the personal agent market fit Openclaw hype refuses to die and while the project had obvious safety issues from day one people still wanted it anyway. Anthropic realized that and just released Cowork's Dispatch feature. I write more about this on the newsletter btw:

love the analogy that Iury gave - think of your AGENTS[.]md like the constitution of your codebase! we also go into an important paper that tries to answer the question that we've all had at some point - are these AGENTS files we're all meticulously creating, even worth it?

This is the real work that teams need to put in, to get the best results from their AI coding agents. listen to our latest episode on Harness Engineering.

Platform engineering for agents! ​If you know plat-eng, you get how tooling helps teams move faster while keeping code standards. Now Harness engineering bridges that concept, turning a repository into an environment purposebuilt for agentic coding. Think we're finally getting out of the weeds!

OpenAI spent 20% of every week cleaning up AI slop from their own coding agents. So they built a "harness" shaping the agent's environment so the slop never ships in the first place. @kau.sh & @iurysouza.dev talk about "Harness Engineering" in the latest ep. fragmentedpodcast.com/episodes/307/

Are AGENTS[.]md files even useful? There’s finally a paper asking the obvious question. @kau.sh and @iurysouza.dev get into the paper, the sync problem, and the small amount of discipline required to stop your agent setup from turning into a mess. 🎧 fragmentedpodcast.com/episodes/306

A System of Record, maps of content and grep. That's pretty much all you need to have HUGE gains in agentic memory. Agents are pretty good at walking text when the path is marked well. This is the main topic in this week's issue: newsletter.iurysouza.dev

Fixing #AISlop with a #SKILL.md in #geminicli Well, at least, removing as much sloppiness as possible from AI generated text! And making it more human-readable ☺️ Inspired by the awesome tropes.fyi Explanations and skill: glaforge.dev/posts/2026/0...

This is why your AI generated text sounds like a lobotomized corpo bot. Either due to time pressure or just basic lack of taste™, a lot of people using AI just don't realize that. I rly hope we can collectively fix this asap.

As a developer, keeping up with AI is super hard and that's precisely why I started this newsletter. These are the best stories I read in the past week. My (human) thoughts on what I think matters and why. No AI slop. 🎧 listen to it on the go :) 🔗 newsletter.iurysouza.dev/p/dead-ides-...

Im not there yet, but I can see this coming.

I started writing a newsletter to keep up with what's going on on AI coding developments. This is the last week's issue recap 🧵. A new one comes out tomorrow. Subscribe! iurysouza.dev/newsletter/t...

Jumped from 20 TPS to 80 TPS ⚡ on my M4 Pro. The trick is that this version uses a MoE architecture with 3 billion active parameters. That makes it way faster. I was running the base model via GGUF (ollama), and now running on MLX (vlm - because this is a vision model too).

The IDE as we used to know is dying and no one bats an eye

If you're using an AGENTS.md generated with /init you're gonna have a bad time.

Friendly reminder that Sonnet 3.5 (on paper it's sonnet 4.5) runs on your Mac for free now. t.co/ZmqCZRRkTC

www.anthropic.com/news/stateme... $ claude --dangerously-skip-constitutional-rights

Function calling everywhere! (kinda looks like MCP though) with AppFunctions If this picks up, the way we use phones will drastically change (and a lot of apps might stop making sense)

Super happy to be back in Paris for @androidmakers.bsky.social! This time I'll be spreading the word on Edge-AI! (Or how to make your phone go brrr!) Now, srsly. Edge AI is heating up (no pun intended) and there's a LOT going on. Watch this space! C'est parti 🇫🇷! tinyurl.com/droidcon-fr-...

First edition of The Feed's Lost and Found is live! I break down harness engineering, code factory loops, and how teams are keeping agent output on track. Fresh out of the press. Get it while it's hot 🔥 iurysouza.beehiiv.com/p/the-feed-s...

Check from the Google office for the Cloud & Al Train-the-Trainer Roadshow in Berlin! Getting knowledge on the more advanced tools and tech for building agents straight from the source. Let's go!

@kau.sh and I bring you some agentic coding goodies on the pod's newsletter. Check it out!

This was fun for both of us! We share some of how the fragmented 🌭 is made + Tips from @iurysouza.dev and me, and a crazy way we used AI to solve very real audio problems 😉.

Big rewrites are coming to a codebase near you. Expect them.

100% this Custom task trackers or planners, systems of records, memory. Also skills specific to each codebases, closing the agentic loop (running tests, linter, check logs, databases, ui components). All of that makes the flow much more reliable and are a must now. mgratzer.com/posts/forgin...

New buzz word: Ever heard of Harness Engineering? At this point, names don't matter that much*, but it's clear the field is changing at neck break speed and the more I read and talk to "practitioners" the more it is clear to me that these will become established practices: 🧵

new episode is out! spot the easter 🥚? recent models have shown massive improvement owing to clever use of "modes" and subagent dispatch. it really clicked for me after chatting with @iurysouza.dev listen to improve your fundamentals! (not just tactics).

🎙️Check out the new @fragmentedpodcast.com episode on subagents! They can be a huge unlock once you understand how it all works and you can get real benefits without having to a swarm them. @kau.sh and I built an RTS mental model for how they work (yes, we did that). Check it out!

I still use bsky, but most of the hot content on AI comes from other apps, newsletters, blogs, and random links. Keeping up is really hard. So I built a personal feed and turned into a public page: iurysouza.dev/feed I'll be also sending a free weekly digest with context if you signup ✌️

👌

If you haven't tried yet, Codex 5.3 xhigh is a really interesting model. (Make sure you select xhigh!) It's really thorough. Takes way longer to answer, but can be really worth it.

CLIs 🤝 Agent Skills It's not because your agent can complete a task with dozens of tool calls and 50k tokens that it should. So what do you do? You can build a script that does the same in a deterministic way in a single 1 milisecond call instead of 4minutes back and forth. 🧵

Skills are an amazing agent tool, but I can't wrap my ahead around people using package managers for installing them. It's like npm supply-chain on roids: - agents can read files - agents can run commands - agents can exfiltrate data - users install them without reading any of the code/text.

Its really useful to get some intuition of how LLMs work so that you can get the most out of them. This was tricky to do, but I guess @kau.sh and I did a good job of breaking down how LLMs work on a 20min, audio-only episode. I always learn smth when I revisit this topic. Check it out!🎙️

As of today, the Agent counter at 39 > Are you not entertained?

It's super interesting to see great engineers going all in on their AI-coding learning journey. This is from Mitchell Hashimoto (from HashiCorp/Terraform, Ghostty): mitchellh.com/writing/my-a... Some things resonate with my experience and some are completely new to me. Great read.

Highly recommend this @gergely.pragmaticengineer.com interview with @booch.com "Software engineering has always been a tale of rising complexity." The abstraction changes, but engineering stays. Industry isn't collapsing, but the hubris won't help either. Be pragmatic!

Just got my second android smart watch ever. 150 EUR off 👌 The last one I had was the og Huawei watch 1.( which still works 7 years later BTW) Let's see if this one holds that long :D

Thinking traces can be so useful! I keep them always on.