
Btw, I have no opinion whatsoever about ChatGPT 6.2 Astra

CTO @ Atherio Join my community:danpatrascu.com/community

Btw, I have no opinion whatsoever about ChatGPT 6.2 Astra


I like the way the wheel is turning. A few months back Anthropic release sub-optimal models while OpenAI did some great imrpovements. All wannabe thought leaders concluded that Anthropic is done. Now it's exactly the other way around

One of the things I've learned working extensively with coding agents is that most of the problems I encounter are actually quite repetitive. Here are five of the most common problems I've encountered and how we address them in my agentic loops: 1️⃣Unsupported claims. Every factual statement the agent makes needs to reference an evidence entry. We have scripts that verify whether the quoted evidence actually exists in the referenced source. 2️⃣Implicit decisions. Every decision that has an impact on the product, scope or architecture is presented to me individually, together with the available options, their trade-offs and a recommendation. 3️⃣Confirmation bias in reviews. Both plans and implementations are reviewed by agents with a fresh context, without being exposed to the reasoning of the agent that produced them. The review loop itself is controlled by a script. 4️⃣False confidence in passing tests. We deliberately introduce mutations into the implementation to verify that our tests can actually detect incorrect behavior. I have recently published in the "Modern Engineering" community all the skills, scripts and verification processes I use in my automated loops. Everything downloadable in a single zip file. If you think it might be useful, you can join and get it here: modern-engineering.circle.so… NOTE: during November you get to join at a 50% discount. You'll keep that discount for as long as you stay subscribed.

I guess I'll be one of the few tech people that will work a lot during the next days and not play GTA VI at all

It's rare that I'm so definitive on something, but if you are an engineer that just started to take AI-assisted engineering seriously a couple of months back, you're already very late. Trying to figure it all out by yourself will just manage the gap and never fully close it. At this point you should be looking into learning from people that are already doing it in production for at least a year.

I have just published in the "Modern Engineering Community" the entire agentic SDLC setup that I use every day while building and expanding Atherio. It contains: ✅17 skills: 10 invokable and 7 skills that are used by the agents themselves ✅12 Python scripts that bring deterministic verification to the agents’ work ✅112 automated checks behind the kit itself ✅A 77-page guide with 14 diagrams, fomr a one-page map down to every single file explained in detail Some highlights that proved to be very useful for me were: ✅a scripted, deterministic system to hold agents accountable to the claims they make ✅a scripted and deterministic mutation testing workflow It's primarily designed for Claude Code but can be adapted to be used with Codex, OpenCode or GitHub Copilot in 5 minutes. And in the community I have also provided the steps to do it. You can download the entire kit by joining the community here: modern-engineering.circle.so…

Is there a way I can stop seeing ads for the Explee scam outreach? I've already blocked 15 accounts yet every day those ads keep coming back from new and new accounts.

Law graduates typically need to practise under an experienced lawyer. Medical graduates go through residency. Architecture graduates typically need to work under an experienced architect before becoming fully qualified themselves. Computer science graduates? They become software engineers out of the box. This will probably change! Yesterday, I wrote about how increasing our aspiration levels when it comes to knowledge and understanding is a must if we want to remain competitive. Today, it's about how I think this will happen. Until AI, one of the main ways engineers gained experience and enhanced their knowledge was through the mechanical repetition of writing and debugging code. As manual code writing and debugging become less frequent, we need to find new ways for junior engineers to gain that experience. And here's where I come back to the intro. I'm not sure if software engineering will ever adopt the formal processes we have for lawyers, doctors or architects, but I'm quite sure that's conceptually the direction software engineering will go. Learning under the guidance of more experienced engineers will become considerably more important than it has been in the past. I also think we'll see learning become increasingly community-based, with engineers sharing their experience, challenging each other's decisions and acquiring knowledge in the context of actual engineering work. That's one of the main reasons I started the "Modern Engineering" community.

One of the things that seem to work incredibly well in my re-designed "factory" is the system designed to catch when the agent makes unverified claims. A lot of hallucination is caught this way:

Idk, to me it feels like collaborating with agents is actually part of the factory itself. Also not sure if the factory itself is something that the model should build. I feel like factories should be quite tailored to each org/team. Not sure generic factories will be a thing
software factories feel like a temporary problem. the models will learn to build them for you the more interesting question to me is how to build a workspace where agents can collaborate with you

People here tend to mistake follower count for expertise. Be careful who you chose to trust! Many influencers are like LLMs: confidently and arrogantly wrong very often.

Seems today I wasn't controversial enough on X 🤣

One of the best parts of my workflow re-design is this: Before automated review starts, we compute hashes of odified code files. An agent then changes the code to introduce bugs. Then we run the tests again and count how many of those bugs were caught by tests. The first 3 implementation workflows introduced 103, 87 and 91 bugs. Bugs that didn't get caught: 7, 4, 2 It's essentially mutation testing without a mutation testing framework.

Cool stuff! I have started one session that will just monitor the planning and implementation sessions for friction and propose improvements to the entire setup while the work is performed.
Today I spent a big part of the day improving my entire SDLC skills, hooks, rules, decision matrices. Curious if this will notably improve my looping.

Agentic coding hack: If you instruct your agent to query your database, ask it first to read the code that writes to the tables it querries.