This is the first article in an @oreilly.bsky.social Radar series on agentic engineering and AI-driven development. The next one gets into what the code actually looks like, and what happened when I tested it against what I intended to build.
www.oreilly.com/radar/the-ac...
My drunken sailor Monte Carlo simulationโthe "Hello, World!" for the systemโwas showing 77.5% probability of falling in the water instead of the theoretical 50%, and I had to pick between three different fixes the AIs proposed. That stuff worries me for people earlier in their careers.
The whole goal of the system I built is to run large-scale simulations and LLM batch workflows, processing thousands of requests through multi-step pipelines with validation and retry logic.
I found the exact opposite โ we need good developers to lead these AIs. I kept catching the AI being confidently wrong in ways that would've been really hard to spot without a few decades of experience.
Either by coincidence or because I'm a genius, I picked the perfect project for testing a structured approach to agentic engineering. (I'm going with genius.)
A big goal for me was cutting through all the "AIs are taking your developer job" hype.
while I handled the architecture and made sure everything actually worked.
The funny part is that the system I built orchestrates AI batch jobs, and the process of building it was... also orchestrating AI.
โฆand see how well it all applies, and what lessons we can learn from a real project. So I ran an experiment: I built a 21,000-line Python system with 900+ tests where I worked with Claude and Gemini for planning and project work, and Claude Code and Cursor did all of the implementation
(Well, technically I ๐๐๐ข๐๐ talk about it, but I didn't want to spoil the surprise!)
I wanted to find out what happens when you let AI write all the code for a real system. And more importantly, I wanted to take everything I've been writing about for 20 years helping developers codeโฆ
๐ ๐ช๐ต๐ฎ๐ ๐ฐ๐ฎ๐ป ๐๐ฒ ๐น๐ฒ๐ฎ๐ฟ๐ป ๐๐ต๐ฒ๐ป ๐๐ฒ ๐น๐ฒ๐ ๐๐ ๐๐ฟ๐ถ๐๐ฒ ๐ฎ๐น๐น ๐๐ต๐ฒ ๐ฐ๐ผ๐ฑ๐ฒ? ๐
๐๐ฉ๐ฆ ๐ฉ๐บ๐ฑ๐ฆ ๐ด๐ข๐บ๐ด ๐ธ๐ฆ'๐ณ๐ฆ ๐ข๐ญ๐ญ ๐ฅ๐ฐ๐ฐ๐ฎ๐ฆ๐ฅ, ๐ฐ๐ณ ๐ต๐ฉ๐ข๐ต ๐ฆ๐ท๐ฆ๐ณ๐บ๐ต๐ฉ๐ช๐ฏ๐จ'๐ด ๐ง๐ช๐ฏ๐ฆ. ๐๐ฆ๐ช๐ต๐ฉ๐ฆ๐ณ ๐ช๐ด ๐ฉ๐ฐ๐ฏ๐ฆ๐ด๐ต.
I've been building something for the last few months and I haven't been able to talk about it until now.
living the dream
Fantastic!! Can't wait to see you there!
Just signed up!!
โ Thursday, January 9th 10AM to 1PM ET / 7AM to 10AM PT
๐ learning.oreilly.com/live-events/...
Really looking forward to this oneโit's going to be great. Hope to see you there!
This is a brand new course, and I'm really excited about it. The course is already over 50% full, so if you're interested, sign up now. Here's the best part: you can take it for free if you sign up for a free trial of O'Reilly Learning.
This is a ๐ก๐๐ง๐๐ฌ-๐จ๐ง ๐๐จ๐ฎ๐ซ๐ฌ๐ where we'll build real code together. You'll learn how to craft prompts that get you maintainable code, write tests that catch AI mistakes, refactor using AI, and most importantly, ๐ ๐๐๐๐๐๐ ๐๐๐ ๐๐๐๐๐๐ ๐๐๐ ๐๐๐๐ ๐๐ ๐๐๐ ๐จ๐ฐ ๐๐๐๐๐๐๐๐๐๐๐.
This Thursday (January 9th), I'm teaching a hands-on @oreilly.bsky.social live event where you'll learn a five-step framework that actually works: context, research, problem framing, critical thinking, and refining.
Illustration showing two developers separated by a gap. On the left, a frustrated developer sits at a laptop with text "Asking the AI for answers." On the right, a confident developer sits cross-legged with a laptop and lightbulb, with text "Working with the AI to solve problems." Header reads "We need to help developers cross the problem-solving gap" with a question mark between the two sides.
๐ ๐๐ฒ๐ฎ๐ฟ๐ป ๐๐ผ ๐๐ฟ๐ถ๐๐ฒ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ ๐ฐ๐ผ๐ฑ๐ฒ ๐๐ถ๐๐ต ๐๐ ๐๐ผ๐ผ๐น๐ (๐๐ต๐ถ๐ ๐ง๐ต๐๐ฟ๐๐ฑ๐ฎ๐!) ๐
Are you looking to get more out of AI coding tools like GitHub Copilot, ChatGPT, or Claude Code? Do you find yourself spending more time debugging AI-generated code than you'd like?
It's another example of what I call the cognitive shortcut paradox. AI tools that make development easier can prevent developers from building the very skills they need to use those tools effectively.
I'd love to know what you think. Have you seen this pattern on your teams?
This is ๐ฑ๐ฎ๐๐ฎ ๐ต๐ผ๐ฎ๐ฟ๐ฑ๐ถ๐ป๐ด. Like physical hoarders who can't throw anything away until their homes become unliveable, data hoarding has the potential to cause serious problems for our teams.
You end up with systems that are expensive to run and impossible to debug. And an entire cohort of developers misses the chance to learn the critical data architecture skills they need to build robust applications.
They know the AI is smart enough to sort through a massive blob of data and pick out the parts that are relevant, so it all just works.
Which, counterintuitively, is actually a problem!
When the AI handles whatever you throw at it, developers stop asking what data actually belongs in the context.
Basically, I've been watching teams adopt MCP (Model Context Protocol) over the past year, and there's a disturbing pattern I keep seeing. A lot of developers like to connect their AI assistants to every data source they can find and then just dump all of that into the context.
๐ ๐ ๐ ๐ป๐ฒ๐ @oreilly.bsky.social ๐ฅ๐ฎ๐ฑ๐ฎ๐ฟ ๐ฝ๐ถ๐ฒ๐ฐ๐ฒ ๐ผ๐ป ๐๐, ๐ ๐๐ฃ, ๐ฎ๐ป๐ฑ ๐๐ฎ๐๐ฎ ๐๐ผ๐ฎ๐ฟ๐ฑ๐ถ๐ป๐ด ๐ท๐๐๐ ๐ฑ๐ฟ๐ผ๐ฝ๐ฝ๐ฒ๐ฑ ๐
I'm really excited about this one! I've been thinking about this problem for a long time, and I think it came out really well.
It's another example of what I call the cognitive shortcut paradox. AI tools that make development easier can prevent developers from building the very skills they need to use those tools effectively.
I'd love to know what you think. Have you seen this pattern on your teams?
This is ๐ฑ๐ฎ๐๐ฎ ๐ต๐ผ๐ฎ๐ฟ๐ฑ๐ถ๐ป๐ด. Like physical hoarders who can't throw anything away until their homes become unliveable, data hoarding has the potential to cause serious problems for our teams.
You end up with systems that are expensive to run and impossible to debug. And an entire cohort of developers misses the chance to learn the critical data architecture skills they need to build robust applications.
They know the AI is smart enough to sort through a massive blob of data and pick out the parts that are relevant, so it all just works.
Which, counterintuitively, is actually a problem!
When the AI handles whatever you throw at it, developers stop asking what data actually belongs in the context.
Basically, I've been watching teams adopt MCP (Model Context Protocol) over the past year, and there's a disturbing pattern I keep seeing. A lot of developers like to connect their AI assistants to every data source they can find and then just dump all of that into the context.
"MCP is a simple but powerful tool with enormous potential for teams. But because it can be a critically important pillar of your entire application architecture, problems you introduce at the MCP level ripple throughout your project." @andrewstellman.bsky.social on #Radar: bit.ly/4rUy0iF
Not sure I agree. I found that (paradoxically) the coding skills we've honed over decades are exactly the skills you need to work with AI. I'd argue that distasteful as you might find it, you're probably better suited to AI assisted coding than almost anybody, at least from a technical perspective.
