Blog

Currently, the use of AI in software development is, at best, disorganized. There’s a maturity curve emerging: companies that are just now learning to use AI tools are running up huge bills as they adopt a variety of AI tools to see which fit their (now-obsolete) workflows. Other companies that have extensively adopted AI, especially in software development, design, and product management, are now laying off thousands of workers deemed obsolete. It feels chaotic. And it doesn’t feel like the chaos will end anytime soon.

But it will. The companies further down the path have visionary leaders who laid the groundwork for agentic engineering. Publish a design in Figma and get a working prototype overnight. This requires a huge investment: workflow orchestration, monitoring, human-in-the-loop at key points in the process, and more. I think of LangGraph and Temporal, and, surprisingly to me, keep reading about companies that have coded their own solutions. Or maybe not surprisingly. After all, AI enables the creation of bespoke software with a single prompt. But the bottom line is that it requires time and money. Lots of it.

How are these companies recouping their investment? Layoffs. Block and Atlassian are the latest companies explicitly saying that AI tools mean smaller engineering teams. It’s not simply that these companies have done the heavy lifting and are cutting workers. They also see the writing on the wall: AI tooling is only going to become more expensive. The cash burn rate for OpenAI ($218 billion through 2030) and Anthropic is astronomical. CFOs know what’s coming: Once investors are done pouring money into AI companies, it’ll be time to collect their returns. Tooling costs will rise. It’s only a matter of time.

That’s why so many tech companies are shedding jobs. Sure, the FAANG companies are cutting jobs to invest in their own AI CapEx. But Block? Vimeo? Wells Fargo? These aren’t companies releasing frontier models. They’re betting that AI tools will help the existing workforce cover the absence of their colleagues. To the tune of 350,000 tech layoffs in the past 15 months.

Which brings me to my point. After the dust settles, we’ll be left with a tech industry with far fewer Software Engineers and far higher AI costs. I recently heard a podcast where the developer of Claude Code compared LLMs to the printing press, democratizing software development. I couldn’t disagree more. Yes, AI tools are here to stay. The chatbots, the coding agents, the orchestration. Everything. But access to those closed systems will become more restrictive and more expensive. It benefits OpenAI to have a chatbot today because it buys public approval at the cost of its investors. It benefits Anthropic to subsidize Claude Code to embed its LLM into developer workflows. But what happens when those products are entrenched and it’s time to turn a profit?

Medieval Europe had more than a handful of printing presses, and they weren’t controlled by a few closed companies that could restrict access at any time. So what does the future hold? Firstly, a division. Companies that can afford it will have access to the latest frontier models, the fully orchestrated workflows, and whatever role Software Engineering is set to become. Others will be running LLMs locally, slowly, and with lower quality output. Not like the old days, but not the gleaming post-scarcity future some enthusiasts are envisioning.

There will be a fast lane and a slow lane. An E-ZPass that allows some workers to zip around, and a toll road for others. Some developers will be using swarms to solve problems, while others chug along on open-weight models as their laptop fans blaze. And I don’t see many people talking about this. It’s due to both the current mood and faith in how the market works – surely if OpenAI and Anthropic turn into premium services, a competitor will emerge?

Maybe. But I don’t see many competitors getting a $200 billion buy-in. Right now I see a proliferation of chaos: new frontier models every few months, execs demanding double or triple productivity boosts, open-weight models releasing seemingly at random, CEOs slashing workforces, security researchers struggling to keep up, and managers and engineers left with a dizzying array of technologies to manage. Things, undoubtedly, will settle. And when they do, and we finally have time to take a breath and look around, I hope we see a million printing presses. Not a few, guarded by white-robed priests holding NFC card readers. Time, and the market, will tell.

I’ve adopted AI. Full stop. I’ve gone through a journey since my last post. At first, it was enough to adopt agentic engineering practices and use AI-assisted development tools to write helper scripts, review PRs, and accelerate my engineering. Then I looked at the things that interrupt or steal time unexpectedly, and tried to automate them. I’ve reached the point where anytime I’m manually interacting with a system – a form, a spreadsheet – I’m wondering why it wasn’t automated in the first place.

What I’m learning is that efficiency in software development using AI tools highlights organizational inefficiencies as engineering bottlenecks are removed. Why are so many company processes manual or Excel-based? Why do I need to read and comprehend an email about the access required for my team to adopt a new tool, instead of the tool and the access being set up for us? Why is there so much process friction that I’m submitting helpdesk tickets to correct configuration errors, request access, or set up new pipelines multiple times per week?

_{Photo by John Cameron on Unsplash}

There’s a debate around essays like “Code has always been the easy part“. Some developers claim that the cost of entry was always high and the need to stay technically current was demanding. I never found that to be the case, professionally. I always found that changing leadership, lack of business requirements, frequent pivots, and organizational inefficiency were the biggest bottlenecks to product delivery. Code was always easy; systems were hard.

So, as I read about engineering groups shrinking and two-pizza teams standardizing around one-pizza teams, I wonder whether the organizations will shrink, too. As a middle manager, I worry about the future. But I also realize that system problems, especially at very large companies, were always my biggest bottleneck as an engineer. If AI helps us collapse bureaucracy so that the people filling out Excel spreadsheets, Google Forms, or IT helpdesk tickets can automate away the need for those systems, we’ll see a massive productivity boost across industries.

Coding was always the easy part. Onboarding is hard. QA is hard. Shipping products is hard. We’ve made the easy part easier. It’s time to focus on the harder parts.