Click Here: Official Java Burn Website
Click Here: Official Java Burn Website
Click Here: Official Java Burn Website
AI generates 10x more code, but your senior engineers still review it by hand and it’s burning them out. Even Google admits code review is now the bottleneck nobody knows how to solve. Florian Buetow, AI engineer at Xebia, has been running experiments to eliminate the human from the review loop entirely, and what he found changes where engineers should focus their effort.
In this episode, we cover:
– Why “stop doing code reviews” is a serious answer (and what replaces them)
– The guardrails that gave the most value: Semgrep rules, architectural unit tests, and stop hooks
– Why your harness matters more than the model
– How Amazon and Google police AI-generated code with policies
– AI burnout, cognitive debt, and “cognitive surrender”: what stays your responsibility
– Step one for adopting agentic software engineering in your team this week
Whether you’re an individual developer drowning in AI-generated PRs or driving AI adoption across a large engineering org, you’ll leave with concrete experiments to run.
More from Florian:
Timestamps:
00:00:00 – Intro
00:00:40 – Code Review Is Software Engineering’s Biggest Bottleneck
00:01:57 – How Amazon and Big Tech Police AI-Generated Code
00:02:55 – Horizontal vs Vertical Scaling of AI Engineering
00:04:37 – Why “No Code Reviews” Might Be the Answer
00:05:22 – Engineering Environments That Give Agents Feedback
00:06:46 – Why the Harness Matters More Than the Model
00:07:21 – When Spec-Driven Development Failed and TDD Worked
00:10:06 – Stop Hooks, Ralph Loops, and Automated Feedback
00:11:30 – The Guardrails That Gave the Most Value
00:14:00 – Architectural Constraints That Keep AI Code Sane
00:15:07 – What Remains a Human Responsibility
00:17:33 – Why All the Hard Work Moves Upfront Now
00:18:47 – The Incredible Skill Junior Engineers Should Learn
00:20:26 – AI Burnout: Why Engineers Are Exhausted
00:22:42 – Cognitive Surrender: Letting the Agent Take Over
00:23:25 – The Hand Grenade Problem with AI at Work
00:24:08 – Outsourcing Code Review to AI Itself
00:26:39 – Teams That Fully Adopted Spec-Driven Development
00:29:01 – Can You Rebuild Software From Tests Alone?
00:30:27 – How to Experiment and Stay Ahead
00:33:15 – Spying on What Subagents Tell Each Other
00:33:59 – Step One: How to Start with Guardrails
00:36:08 – Data Mining Your Session Logs for Patterns
00:37:00 – Stuck With One Harness? Here’s What to Do
00:38:28 – The One Experiment to Run This Week
#softwareengineering #aicoding #codereview
Connect with Patrick Akil:
Click Here: Official Java Burn Website
Click Here: Official Java Burn Website
We need to hear more about the daily struggles from people in the trenches. There are some very serious and terrible trends going around that are profoundly affecting peoples work lives. For instance, at the extreme, the narrative of “AI can replace your developer teams” is taken by some as an instruction to give 50% of a team AI and fire the rest. Anyone using these tools daily will tell you it’s not that simple. The best part is, if you’re an engineer, these emerging problems are ENGINEERING problems, and can be challenging and, dare i say fun? Great to hear from someone who is working on the edge of what is possible. Highly recommend watching
You have an excellent interview skill. Your questions are insightful and you clearly listen and internalize and respond. Very impressed with your format and clear preparation. It develops good and valuable conversation with a variety of perspectives and I feel like you try to extract the insights and opinions of the guest rather than project your own viewpoints. Great job. I've subscribed.
Great conversation, I love hearing from actual engineers building and releasing code. There is so much noise from tech bro's posting how they 100X'd with agent flows, but never actually show or go into details about anything.
Building real software, within a team using agentic flows, is the real conversation.
If ai still drifts and still hallucinates how can we trust AI generated PR reviews . We need a deterministic approach
The "cognitive surrender" part is so true. A few months in, you own code you never really understood. Guardrails keep new code clean, but they don't tell you why the agent did it that way. I keep a simple "how it works" note with the reason behind each change. It's the only thing that makes big changes not scary.
I didn't watch the entire thing, but Code Review MUST be done, religiously today in the AI era more than ever. HARD NO
How convenient that the guy that works at an ai company has identified a bottleneck caused by ai and his solution is to just skip that bit
if your shit stinks, stop shitting…
Code is context, so code quality matters for the LLM. That was insightful.
It only works if your code doesn't matter…. Good luck explaining this process to the FDA….
So you solved one bottleneck to just create a new bottleneck! Engineering at is best.
I find it quite inefficient, but actually good fun
This ain't my YouTube account but one thing I would like someone to experiment with and the free time is making a program for termites or any CLI software will there be powershell or CMD or like a native AI within power shower CMD within the windows OS itself I can take any Chromebook or PC and then reverse engineer Windows 11 specifically for that hardware set making it into a real native Windows PC that's something I've been experimenting with myself and also building a real bottom of a I like the one seen in the fictional series 3rd season of sword art online and because it works differently than a top-down AI model and being that has human consciousness and then Free Will and the ability to think for itself has more behavioral aspects of it I can really affect coding itself in major ways like also being able to have its own biased towards coding I've been experimenting with that I was able to code one in termux using quad code in an llm and then I train the llm to build the architecture to build the app in the app didn't really turn out the way I wanted it so I had to take all the files that I created for the bottom of AI and applied it to itself and ran it through time dilated training sessions no accelerated time dilation training sessions training itself in years and years and years of coding in just minutes and different coding different coding units and then not only that I did it I also trained it in specific coding aspects as well before applying the files to it and it was really amazing to see
LLMs produce random text/code. So in order to verify what is going on, you need code reviews. You can try guardrails, but they can be overcome. You can try to skip code reviews and hope for the best. Good luck.
I’m a new grad so maybe i just dont know what im talking about, but all this seems like more work than just programming it yourself + I love programming, I don’t want to outsource the craft to some llm. I wish I was born like 25 years ago so I could experience the golden years of software development
Removing human code reviews loses the only mechanism that ever had a chance of catching what nobody knew to look for.
You should put the guest's name and some credentials in the video title and description. The first thing I want to know is who I’m listening to. In this case, the guest is working for an AI consulting company and has a vested interest in having as many people as possible hire them for consultancy. That’s not a neutral perspective, and it should be prominently displayed.
I think it is bad honestly .
man this is what i needed to hear
i never thought about it this way 🤔
finally someone gets it, so true