Why Software Factories Fail

"Read the Code!" is one of the key ideas from Dex Horthy's talk "Why Software Factories Fail".

The author touches on a very hot topic right now: we are actively pushed to put AI-generated code into production, build software factories, and eliminate the human bottleneck from the process. And all of that would be great if, at the same time, the overall quality of software products wasn't going down and the number of incidents wasn't going up.

Dex explains this phenomenon by pointing to a fundamental limitation of current models: they can't maintain and improve the quality of a codebase over time. Models are trained to solve the issue and pass the tests. There is nothing there about good architecture, clean code, or future maintainability.

If the model knew what good code looks like it would write it in the first place.

What to do with that? Dex suggests to do the following:
🔸 Make Product Review: align goals, check mockups, etc.
🔸 Do System Architecture: establish clear component boundaries and API contracts.
🔸 Perform Program Design: define types and method signatures, test approach, program layout, and dependency stack.
🔸 Implement by Vertical Slices: build features vertically, reducing the scope of what an agent can change in a single iteration.

So, it's all about shifting review left to earlier stages and focusing more on architecture and program design (I described similar ideas about code review here). And then let agents make small changes that engineers can actually understand and review.

And yes, we still need to review the changes.

#engineering #codereview #ai

Tracer Bullets

Continuing the topic from the previous post, let's talk in more detail about implementing vertical slices.

The original concept comes from The Pragmatic Programmer, where it was called tracer bullets. A tracer bullet is a small, end-to-end slice of functionality that touches all layers of the system at once.

Matt Pocock in his article adapts this concept to AI engineering:
- Build a small feature end-to-end
- Test it immediately
- Get feedback
- Move to the next slice in a fresh context window
- Repeat

Matt explained this in the following way: AI's natural inclination is to build big layers in isolation. But we need to do the opposite: build the whole scenario end to end across all layers.

The article also provides a prompt sample:

When building features, build a tiny, end-to-end slice of the feature first, seek feedback, then expand out from there. When building systems, you want to write code that gets you feedback as quickly as possible. Tracer bullets are small slices of functionality that go through all layers of the system, allowing you to test and validate your approach early. This helps in identifying potential issues and ensures that the overall architecture is sound before investing significant time in development.

So this approach forces the agent to think by small chunks and produce more predictable results.

As they say, everything new is just well-forgotten old. As we can see, AI often doesn't bring really new concepts, but rather brings existing ones back to life.

#engineering #ai

AI & Repository Strategy

For many years, there has been an ongoing debate between monorepo and polyrepo fans (I'm on the monorepo side, by the way, but that's not the topic today). Today, I'd like to look at these approaches from a slightly different angle: the speed of introducing changes.

Over the past year, our industry has gone through a significant transformation. We changed harnesses, introduced skills, built agents, engineered loops, adopted SDD, and so on. And it seems we're still far from the end.

All of this requires a very high speed of adoption and adaptation at the team level. This means that everyone should use the same SDD framework and approach to specs, the same set of skills and their versions, and the same agents built into the workflow. Across all repositories in use.

And that's where I see a huge difference.

With a monorepo, you just need to bring the required set of skills or CI steps into the repository, commit them, and that's it. Whether the team wants it or not, the agent starts using the required skills, the required automation steps start running for everyone, and the team learns faster because they can't bypass the new rules committed to the repository.

With polyrepos, everything is much more complicated (the problems start with tens of repos and become much more complicated with hundreds). Not only is the context fragmented and needs to be assembled into a virtual monorepo, but you also need to bring the same set of skills into all repositories, control the process of updating them, figure out where and how to store SDD specs, and deal with other problems.
From what I see in practice, everyone invents their own wheel to somehow solve this problem. But it takes time and resources. It's slow.

So it turns out that when it comes to the speed of introducing changes across a team, monorepos definitely win.

#ai #engineering #ci #leadership

Fixing the PR Bottleneck

Pull requests have always been a bottleneck, and today the situation has become even worse. Agents generate changes much faster than humans can review them.
Let's check what Matt Pocock recommends doing about that in his new talk Fixing the PR Bottleneck.

First of all, he suggests using a three-layer code review approach:
1. Automated checks: linting, type checking, testing.
2. Automated review: agents that enforce coding standards.
3. Human review.

The main goal is to make human review faster by relying on the first two layers. And to make automated reviews effective, you need to teach your agents your codebase design rules: how the code is structured, how to organize modules, how to write tests properly, etc.

Some recommendations on how to organize the process:
🔸 Keep codebase design rules separate. Describe them in a CODING_STANDARD.md file, not in AGENTS.md, and run the review in a separate session.
🔸 Don't outsource automated code review. Generic code review doesn't work well in practice. It produces too many false positives.
🔸 Automated reviewers should fix issues themselves. Agent should not create review comments on PRs and expect human to read them, it should provide fixes before human review.
🔸 Some PRs are more important than others. Categorize them by blast radius and include this information in the PR.
🔸 Run retrospectives regularly. You shouldn't have to leave the same comment twice. Teach your agents and incorporate this knowledge into your coding standards.

There is definitely a lot to think about how modern code review should be organized.

I personally liked the idea of having a code review agent fix issues instead of leaving comments for humans to handle. Anyway an engineer will just ask another agent to fix them first. So there is no need for additional human intervention here. I'll definitely try this in my teams.

Also, Matt has a pretty popular GitHub repository with skills. There are a lot of useful engineering skills there that you can try out with your own team.

#ai #engineering #codereview

About Presentations

Last month, I prepared two presentations for public talks. And I also regularly prepare presentations internally at work.

And here's what I've noticed:
🔸 You can recognize a Claude-generated presentation from a mile away. All those rectangles with tiny text inside. Once you notice it, you can't unsee it.
🔸 AI presentations are too literal. They explain everything and leave no intrigue for the audience. It's more of a document than a real presentation.
🔸 AI presentations are incredibly boring. Seriously.

If you weren't good at making presentations before, AI can give you a great boost. But if you were good at it, AI still can't produce a comparable result.

A presentation isn't just about data. It's storytelling, intrigue, and drama. A presentation should help capture and keep the audience's attention. And that requires developing communication and presentation skills. I mean, human skills :)

Of course, I use AI to prepare presentations. But for important talks, I mostly use the result as a draft and a quick start. And I still spend quite a lot of time making it actually interesting and worth watching.

To sum up, my rule is pretty simple:
If you need a presentation just to tick a box or for reporting purposes, you can safely leave the whole thing to AI.
But if you want people to remember your talk, get interested, and actually pay attention, you'll still have to put some effort into it.

#softskills #presentationskills #ai

Software Factories

We haven't even had enough time to fully adopt SDD or build deep expertise in loop engineering, and they're already no longer the hottest topics. Now there's a new concept on the hype wave: Software Factory.

Software Factory is a software development approach where autonomous agents handle implementation, testing, and validation within standardized workflows, while engineers define requirements, architecture, quality standards, and guardrails.

The idea is that we should move from just implementation (inner) loops to building factories that orchestrate those loops (outer loop). In other words, we should create a system that can build a software on its own.

Factories can be:
🔸 Light factories: Humans are still in the loop.
🔸 Dark factories: Agents define scope, build, and ship code without anyone really reading the details.

I would say only some startups are brave enough to use dark factories (and it's also fine for pet projects). Companies where mistakes can have a real impact still prefer to keep humans in the loop.

But by the way, code generation has never been the hardest part of software engineering. The most complex part has always been understanding and agreeing on what we actually need to build, how to build it, and how to verify that we built the right thing with the expected quality.

Agents just make it more visible.

For a deeper dive:
- Addy Osmany Software Factories, Light and Dark
- Zach Lloyd Software Engineering Is Becoming Factory Engineering
- Tereza Tížková What It Actually Takes to Build a Software Factory

#engineering #ai #ai4sdlc

Skills Distribution

Nowadays, skills are everywhere... Thousands of them have been written, some useful, some not so much. But the questions of proper management and distribution are still not solved, especially at scale.

Imagine a medium-sized engineering department of 50 people: 6 separate teams, 20+ applications, and 50+ repositories.
Now you start introducing SDD, shared skills for development, documentation updates, code reviews, etc. And here comes the famous human factor: someone installed the skills, someone forgot, someone installed only a few, someone updated them, and someone just didn't bother.
As a result, you have no idea who uses what, and predictable results are out of the question.

My position is pretty simple: anything that requires even occasional manual actions to keep working will never work.

The solution I chose for my team is quite straightforward:
🔸 Required internal skills are declared in the root apm.yml file (you can read about APM here).
🔸 The skills are committed directly to the repository for all required agent harnesses.
🔸 A nightly CI job checks apm.yml for updates to the installed skills and automatically commits any updates.
🔸 Profit!

The same approach can be used for external skills and npx packages, but that would also require security checks for vulnerabilities and prompt injections. We must be much more careful with external skills.

So, what does this give me?
- The team uses the same set of basic development skills.
- The team uses up-to-date versions of those skills.
- I vibe coded a report that shows if the required skills are present in the repository and the CI job is configured.

Maybe it's not the most elegant solution. We intentionally duplicate skills across repositories. BUT! It works, and it works reliably. Moreover, it costs almost nothing in terms of internal infrastructure and maintenance.

#engineering #ai #ci

Guns, Germs and Steel

Guns, Germs, and Steel: The Fates of Human Societies is a popular science book by American scientist Jared Diamond. The book tries to explain why some civilizations are more successful than others.

To answer this question, the author goes back to ancient history, particularly to the domestication of plants and animals.
He explores how domestication actually works, what conditions are required, and how humans have influenced the evolution and selection of plants and animals. For example, wild strawberries and modern cultivated strawberries look very different. And there are hundreds of examples like this.

Interestingly, not all animals are suitable for domestication. For example, horses were domesticated, but zebras weren't. According to the author, zebras simply have a terrible personality and bite too much 😃.

And what conclusion does he draw from all this?
One of the most important steps in the development of civilization was the transition from a nomadic to a settled lifestyle. And this transition was only possible when suitable plants and animals were available for domestication.

Another advantage was the east-west orientation of a continent. Similar climates made it easier to exchange crops, livestock, technologies, and diseases.
Eurasia stretches from east to west, while Africa and the Americas stretch mostly from north to south. According to Diamond, this is one of the main reasons why civilizations in Africa and the Americas developed differently and were eventually colonized. And during colonization, Europeans also brought plenty of unfamiliar diseases, which devastated local populations.

Overall, the book is quite interesting, and I know people who absolutely love it.
Unfortunately, I'm not one of them.
Despite the interesting concept, I found the book incredibly tedious. It was really hard for me to finish. I don't think I've ever read so much about agricultural crops in my entire life. Apparently, agriculture is just not my thing 🙃.

#booknook #science

AI Coding Autonomy Levels

A few days ago, I read the early release of Addy Osmani's Agentic Engineering ( about that in the next post). And it contains a few references to Dan Shapiro's The Five Levels of AI Coding Autonomy.

Since this concept has become quite popular, I think it would be useful to put it there.

So the main goal of the model is to establish a common language for where we are and where we're going in AI-assisted software development:
🔸 Level 0: Spicy Autocomplete. AI is used as a reference tool or smart search. The developer writes code manually with AI suggestions.
🔸 Level 1: Coding Intern. The developer prompts AI to write specific functions, then reviews and integrates the output.
🔸 Level 2: Junior Developer. Pair programming with AI, where the human reviews the generated code in real time.
🔸 Level 3: Developer. AI generates most of the codebase. The human's role is mostly to manage the process and review code. That's the famous human in the loop.
🔸 Level 4: Engineering Team. AI handles complex tasks without human supervision. Humans rely on the system's own checks and validations.
🔸 Level 5: Dark Software Factory. The engineer manages the system, not the code. It's a black box that turns specs into working software.

Most engineering teams are currently somewhere around Level 2. But Level 5 looks like the future we're moving toward.

#ai #engineering

Agentic Engineering (Book)

Addy Osmani is a well-known AI evangelist. He spent 14 years at Google and is now a member of the technical staff at Anthropic. Right now, he is writing a book called Agentic Engineering, which is already available as an early release on the O'Reilly platform. The planned release date is September 2027, and only five chapters are available for now.

So what's inside:
🔸 Accountability. What is acceptable for a pet project may not be acceptable at work. We are accountable to our customers for the quality and the results.
🔸 Engineering discipline. AI adoption requires strict engineering discipline: "Grant an agent the autonomy you can check, per task."
🔸 Key components of AI development: agents, context, harnesses, hooks, loops, etc.
🔸 Judgment is still a human responsibility. Models are not trained to maintain code maintainability. Dex Horthy explains the same problem in his talk "Why Software Factories Fail," which I described here.
🔸 Context engineering. Main principles, challenges, and why it's so important to put the right things into it.
🔸 AGENTS.md is a living artifact. The author highlights that if you simply generate your AGENTS.md with AI, you don't really provide anything useful to the agent. AGENTS.md should contain project-specific knowledge and lessons learned, but it shouldn't become too large: outdated instructions should be regularly removed.
🔸 Importance of specs. Developers must carefully prepare and review the specs. Everything you don't describe in enough detail, the agent will figure out on its own. And you probably won't like the result 🙃.

I really enjoyed the book. It's easy to read, and instead of trying to hype AI, he describes everything from an engineering perspective, honestly explaining the risks and blast radius of different decisions.

The book covers fundamental principles of AI usage and provides practical recommendations on what to use and how to use it. While reading, I got several ideas I could apply in my own work. For example, I've already planned an audit of AGENTS.md files across my teams.

As the industry is changing very fast, I recommend reading it now rather than waiting for the final release. Even these five chapters contain a lot of useful recommendations (and, of course, I'm waiting for the new chapters).

#booknook #ai #engineering