Thoughtworks CTO: AI Means We Need Developers More Than Ever
“I saw something on X where somebody had basically said, ‘Look, I vibe coded this app. I don’t need developers. I’m cool.’ And then, within 24 hours, they’d been attacked. It had not all the security vectors covered, or any of the stuff that you need to think about when you know you’re a deep software developer.”
This is how Rachel Laycock, CTO of Thoughtworks, kicked off a very reflective conversation with The New Stack on the future of AI and its current impact on the tech industry.
Soon after, she said that the X user wrote: “I didn’t know what I was doing, but the people who were attacking me clearly do.”
Laycock is not surprised because, after all, these AI tools are trained on the internet.
“The internet isn’t necessarily full of great code and generating more of that is not necessarily good for us,” she explained. “There’s this massive focus on productivity and creating code as fast as possible. Because backlogs are endless, and everybody’s complaining that the IT or the tech department’s not fast enough, and we need to get more features out.”
But, to her the biggest challenge in the tech industry is legacy code — which is only going to get worse with AI-generated code at scale.
AI does have a role to play in paying down that debt and finally moving to the cloud, but Laycock argues that only increases the demand for deep-thinking, problem-solving engineers. Read highlights from this exclusive conversation for how AI can augment the software development workforce, not replace it.
AI and Legacy Modernization
A lot of organizations are betting that AI agents are going to get smarter and then better. A lot of hope is particularly being poured into retrieval-augmented generation or RAG-based AI to try to improve the models and tools even faster.
But we aren’t near understanding the long-term impact of AI.
“It’s not clear when we will get to a settled space of what are the top three or top five tools and models people use for this. It’s very, very changeable,” Laycock said. Add to that, “most of the stuff that people are demonstrating is building a Greenfield app,” which, she said, is relatively easy.
Legacy modernization is still the top challenge facing most enterprises.
The current market, she continued, is “focusing too much on efficiency of producing code, which isn’t actually the problem in the industry.”
Overall, legacy modernization is still the top challenge facing most enterprises. On top of that is the knowledge fragmentation across hundreds of apps that further hinders this move to the cloud.
This echoes true with what developers complain is slowing them down: technical debt and documentation.
Both business and tech departments are stuck in trying to understand code. Adding only more AI-generated code decreases that comprehension more.
The True Cost of Trying to Replace Developers with AI
“Meanwhile, everybody’s like: How can I just do this as fast as possible? I don’t want to hire anybody new. I want less developers,” Laycock said.
The industry is banking on AI job replacement before it’s even proven at scale yet. And after malicious npm packages found infecting Cursor AI with backdoor steal credentials, it’s also not clear that security is in place at scale either.
“From an agentic AI perspective, what I’m seeing works is things that are more focused on a specific task,” she said. “So the idea of like ‘Fix this Jira ticket,’ but don’t say, ‘Fix these 100 tickets,’ because it will go off in probably many endless loops and cost you a fortune in tokens.”
This isn’t like GitHub Copilot, which has a nominal cost of $100 per developer per year. Estimates for the cost of AI agents can run up to potentially tens of thousands per developer per year. So-called “token-omics” will definitely be the next FinOps priority.
There’s a theory that an AI-driven supervising model will be cheaper than paying software development salaries. Except that, again, no one has proven the capability of the AI at scale, nor have they run the numbers on the true cost.
“If I had lots of agents running discrete tasks — basically an agent farm — and then you had senior developers (or even junior developers, which again sounds dangerous) basically supervising them,” Laycock said. “You’ll hear from product companies or from the big cloud service providers who are incentivized to get everyone to use lots and lots of tokens.”
“Nobody’s talking about what [AI] can’t do and what things need to be solved for; and generating more code is not actually going to help us.”
– Rachel Laycock, CTO of Thoughtworks
“If Microsoft, Google and AWS are all saying the same thing. Well, they want people using GPUs, right? That means token usage” she continued. “They want you to build agents running around. They don’t care if they go in endless loop. Or maybe they do care, but they’re incentives to get workloads, right?”
This risk doesn’t mean rejecting AI completely. Laycock finds some more experienced peers simply dismissing it — which is its own kind of risk.
“We need to find some kind of middle ground of: What can it do? What can we use this for?” she said. “Don’t just dismiss it. Don’t use it in anger. These are going to keep changing and improving, but it’s your experience that will tell us where the gaps are and that’s what I feel like is missing at the moment. Nobody’s talking about what it can’t do and what things need to be solved for; and generating more code is not actually going to help us.”
Right now, it’s about getting teams to adopt these new AI tools to test boundaries, so that more senior engineers, with 10 or 20 years of experience, can help solve those boundary conditions. And of course have those senior engineers train the next generation of junior engineers — because you cannot have senior engineers without them being juniors first.
For Thoughtworks’ mostly enterprise customer base, Laycock said, there’s a sense of conservatism — they’re “waiting to see where things land.”
“Because if you think about rolling out a change, even rolling out GitHub Copilot is big for them at scale,” she said. “With tens of thousands of developers, it’s not an easy change.”
When you are rolling out at that scale, these enterprises need to be quite sure they are betting on the right tooling and model choices.
She continued, “The whole landscape is not settled enough for them to do that at scale.”
CodeConcise: Overcoming the Mystery of Legacy Code
One of the main barriers to digital modernization and the move to the cloud is simply that less and less people are around who actually built these older systems.
Teams, with incomplete documentation and architectural decision records, are left uncertain what’s truly a zombie service and what’s one that the whole business relies on.
AI — great at explaining complexity — is part of this solution. But, as with Big Bang modernization approaches, it’s nowhere near the whole part of the solution. Nor can it be just using AI to regenerate code, Laycock said, because you’ve got to understand the mainframe code to be able to translate it to the cloud native context.
“One of the challenges we face is that whole understanding the codebase: Do we know what it’s doing?” she said. “Why it’s doing it?”
Thoughtworks has built a new generative AI tool, CodeConcise Legacy Assistant, which indexes code along with a context window and a conversational AI overlay, to help clients understand their systems. With CodeConcise, the Thoughtworks team aims to build a context window in partnership with subject matter experts.
By no means does Thoughtworks think it’s unique in this approach to solving the ever-present move-to-the-cloud problem.
Instead of a Big Bang lift and shift approach, Thoughtworks advocates using AI to help understand your systems so you can make the right cuts across dependencies and teams, working with everyone to then modernize that cross-functional piece.
“People are thinking about how you can use AI to basically understand the workflows and the data flows, and then translate — basically rebuilding the codebase,” she said. It could be a year or even 10 years away, but “the idea that you could use AI to regenerate the code is a mind-blowing thing, if you think about some of the problems we’re trying to solve. We aren’t there yet, but you can start doing things with that.”
It is the time for early AI experiments pursuant to those later game-changers.
“Understanding something big, complex, somewhat structured, somewhat unstructured, in different languages, being able to interrogate it,” mused Laycock. “So we built a chat interface that says: What does this do? Who are the users of the system? How are they using it? Explain this functionality. Things like that.”
Instead of a Big Bang lift and shift approach, Thoughtworks advocates using AI to help understand your systems so you can make the right cuts across dependencies and teams, working with everyone to then modernize that cross-functional piece.
“We’ve always taken what we call a thin slice,” which she explained is “an interactive approach to identifying what are the different domains, contexts and models that you can leverage that helps you slice up and create seams around pieces you need to change, and move forward to the process [of] identifying dead code and things that you don’t need.”
Wanted: More Developers for AI Experimentation
While this dream-to-reality is still in the works, Laycock sees the industry stuck on the code generation piece, when solving the legacy problem is a more interesting and impactful Gen AI use case for most enterprises.
“Because that’s the reason they can’t move faster,” she continued: the legacy applications and the legacy data structures don’t support fine-tuning models. AI will also have a role to play, she predicts, in adapting the data architecture to be able to support building AI applications.
“Those are two big, not straightforward problems that AI, I think, can augment and support in,” Laycock predicted, “but it’s not just going to magically you press a button.”
For now, Thoughtworks is crowdsourcing a backlog of hypotheses of where the tech industry will head next. From app migration and modernization to code generation to AI agents and beyond, Laycock says each organization has to take this product thinking approach, testing out theories, examining and testing AI across the software development lifecycle, not just that inner loop code writing bit.
“AI is being infused in all these things, which is great, but we have to remember it’s not deterministic.”
– Laycock
“Now they’re just like: How can we make the bit at the start as fast as possible and not need developers? And I’m like: Why does it always come back to just not needing developers?” Laycock said.
“I get it. We’ve gone through various talent wars. It’s hard to hire good people, but you’re kind of scaring away people even [from] joining the industry, and we haven’t proved that we don’t need them yet.”
The focus shouldn’t be eliminating the tech talent, she argues, but on leveraging AI to augment their roles. This can be through abstraction, eliminating toil, generating tests, writing docs and capturing that holistic view of the developer experience.
In this fundamental shift, Laycock said, don’t forget to observe everything as you put AI-generated everything or anything into production.
“Observe it. AI is being infused in all these things, which is great, but we have to remember it’s not deterministic. And there’s probably a lot more training and tuning that needs to be done on some of these more complex problems before we’ll get to everybody’s Holy Grail,” Laycock said.
In the end, she said, “I am convinced that people’s jobs will change, but [as to] how work will change and is changing,” as an industry we’re nowhere near having figured that out yet.
