Introduction: Thoughtworks CTO believes that while artificial intelligence is driving a fundamental shift in the tech industry, we must retain the work of developers.
“I saw someone on X saying, ‘Look, I feel good writing this application. I don’t need developers anymore, I think it’s cool.’ However, less than 24 hours later, their software was attacked. It didn’t include even the most basic security considerations, nor did it take into account any factors that a deep developer needs to think about.”
This was a very profound conversation initiated by Rachel Laycock, Chief Technology Officer at Thoughtworks, on the future of artificial intelligence and its current impact on the tech industry.
Shortly before, she commented on the X user: “He didn’t know what he was doing, but the person who attacked the application was clearly knowledgeable.”
Laycock said she wasn’t surprised, as after all, these AI tools are trained on the internet.
“The internet is not necessarily full of good code, and generating more of this code may not be good for us,” she explained. “Everyone is now very focused on productivity, hoping to write code as quickly as possible. Because the backlog of work is endless, everyone is complaining that the IT or technology department is not fast enough, and we need to launch more features.”
But for her, the biggest challenge facing the tech industry is legacy code – and with AI-generated code scaling, legacy code will only get worse.
AI is indeed playing a role in paying down technical debt and eventually migrating to the cloud. But Laycock believes this will only increase the demand for engineers with deep thinking and problem-solving skills.
AI and Legacy System Modernization
Many organizations are betting that AI agents will become smarter and better. There are high hopes for Retrieval Augmented Generation (RAG) or RAG-based AI to improve models and tools faster.
But we are still far from understanding the long-term impact of AI.
“It’s not yet clear when we will be able to determine the top three or five tools and models people are using in this area. The situation is changing very, very much,” Laycock said. Furthermore, “most of what people demonstrate is building a Greenfield-like application,” she said, which is relatively easy.
Modernization of legacy systems remains the biggest challenge for most enterprises.
She continued, the current market is “too focused on the efficiency of code generation, but this is actually not the industry’s problem.”
Overall, the modernization of legacy systems remains the biggest challenge for most enterprises. In addition, the fragmentation of knowledge across hundreds of applications further hinders enterprises from migrating to the cloud.
This echoes the factors developers complain about slowing them down: technical debt and technical documentation.
Both business and technology departments are struggling to understand code. The more AI-generated code there is, the worse the understanding becomes.
The Real Cost of Trying to Replace Developers with AI
“Meanwhile, everyone is thinking: How can I get this done as quickly as possible? I don’t want to hire any new people. I want to reduce the number of developers,” Laycock said.
The industry is hoping that AI will replace existing jobs even before its large-scale application is verified. After discovering that a malicious npm package infected Cursor AI with a backdoor for stealing credentials, it is unclear whether its security is sufficient for large-scale adoption.
“From the perspective of AI agents, the things I’ve seen that are effective are more focused on specific tasks,” she said. “So, for example, ‘fix this Jira ticket,’ don’t say ‘fix these 100 tickets,’ because this could get stuck in countless infinite loops, costing you a lot of tokens.”
This is unlike GitHub Copilot, which has a nominal cost of $100 per developer per year. It is estimated that the cost of AI agents could be tens of thousands of dollars per developer per year.
There is a theory that an AI-driven oversight model is cheaper than paying software developers’ salaries. However, similarly, no one has proven the scalability of AI, nor has anyone calculated its true cost.
“If I have a lot of agents running dispersed tasks – basically an agent farm – and then you have senior developers (or even junior developers, which sounds dangerous) supervising them,” Laycock said. “You will hear product companies or large cloud service providers say they are motivated for everyone to use a lot of tokens.”
“No one is talking about what AI cannot do, and what problems need to be solved; and generating more code actually doesn’t help us.”
—— Rachel Laycock, Chief Technology Officer, Thoughtworks
“If Microsoft, Google, and AWS are all saying the same thing. Well, they want people to use GPUs, right? That means using tokens,” she continued. “They want you to build agents that run. They don’t care if these agents get stuck in infinite loops. Or maybe they do care, but the incentive they offer is to get the workload, right?”
This risk doesn’t mean completely rejecting AI. Laycock finds that some more experienced colleagues simply ignore it, which is itself a risk.
“We need to find some middle ground: What can it do? What can we do with it?” she said. “Don’t dismiss it easily. Don’t use it in a fit of anger. These features will continue to improve, but only your experience can tell us where the gaps are, and that’s what I currently feel is missing. No one is discussing what it cannot do, and what aspects need to be addressed; and, generating more code actually doesn’t help us at all.”
Currently, the key is for teams to adopt these new AI tools to test boundary conditions so that more senior engineers with 10 or 20 years of experience can help address these boundary conditions. Of course, also letting these senior engineers train the next generation of junior engineers – because without the experience of junior engineers, it’s impossible to cultivate senior engineers.
Laycock said that for Thoughtworks’ main enterprise client base, there is a conservative sentiment – they are “waiting to see how things develop.”
“Because if you want to roll out a change, even just rolling out GitHub Copilot is a big deal for them,” she said. “For thousands of developers, this is not an easy change.”
When you roll out at this scale, these enterprises need to be very sure they have chosen the right tools and models.
She continued: “The whole situation is not stable enough for them to do this on a large scale.”
CodeConcise: Unlocking the Mystery of Legacy Code
One of the major obstacles to digital modernization and migration to the cloud is that fewer and fewer people who actually built these old systems remain.
Due to incomplete documentation and architectural decision records, teams cannot determine which are truly zombie services and which are services that the entire business relies on.
And AI is good at explaining complexity, it is part of this solution. But, like the “big bang” modernization approach, it is far from the whole solution. Laycock says it is also not possible to simply regenerate code using AI because you have to understand mainframe code to convert it to a cloud-native environment.
“One of the challenges we face is understanding the codebase holistically: do we know what it’s doing? Why is it doing it?” Laycock said.
Thoughtworks is building a new generative AI tool, CodeConcise Legacy Assistant, that can index code and comes with a context window and conversational AI overlay to help customers understand their systems. With CodeConcise, the Thoughtworks team aims to work with industry experts to build a context window.
And Thoughtworks doesn’t believe its approach to solving the pervasive problem of migrating to the cloud is unique.
Thoughtworks does not advocate a “big bang” lift-and-shift approach, but rather using AI to help understand your systems so that dependencies and teams can be properly partitioned, and work with everyone to modernize cross-functional parts.
“People are thinking about how to use AI to understand workflows and data flows, and then transform – which is rebuilding the codebase,” Laycock said.
This could take anywhere from a year to a decade, she said: “Thinking about some of the problems we’re trying to solve, the idea of using AI to regenerate code is really exciting. We’re not there yet, but you can start using AI to do some things.”
Now is the time for early AI experiments based on those game-changers that will come later.
“Understanding something large, complex, structured or unstructured, analyzed in different languages, and being able to explore it,” Laycock said thoughtfully. “So we built a chat interface where you can ask: What does this system do? Who are the users of the system? How do they use it? Explain this feature. Things like that.”
Thoughtworks does not advocate a “big bang” lift-and-shift approach, but rather advocates using AI to help understand your systems so that dependencies and teams can be properly partitioned, and work with everyone to modernize cross-functional parts.
“We have always taken what’s called the slicing method,” she explained, “which is an interactive approach to identify different areas, contexts, and models that can be leveraged to help you slice and create seams for the parts that need to be changed, and continue the process of identifying dead code and what is not needed.”
Hiring: More Developers for AI Experiments
While this dream is still in the making, Laycock believes that the industry is still stuck in the code generation phase, whereas solving legacy issues is a more interesting and impactful generative AI use case for most enterprises.
“Because that’s why they can’t move fast,” she continued: legacy applications and legacy data structures don’t support fine-tuned models. She predicts that AI will also play a role in adjusting data architecture to support building AI applications.
“I think these two problems are ones that AI can enhance and support in solving, they are not simple problems,” Laycock predicted, “but it won’t solve them magically like pressing a button.”
Currently, Thoughtworks is using crowdsourcing to gather a series of hypotheses about the future direction of the tech industry.
From application migration and modernization, to code generation, to AI agents and more, Laycock says every organization must adopt this product mindset to test theories, check, and test AI throughout the software development lifecycle, not just the internal loop of code writing.
“AI is integrating into all of these things, which is great, but we have to remember that it’s not deterministic.”
—— Laycock
“Now they just think: How can we get the initial part done as quickly as possible without developers? And I think: Why always come back to the problem of not needing developers?” Laycock said.
“I hope you understand. We’ve been through various talent wars. It’s hard to hire good talent, but doing this scares away some people, even making them afraid to join the industry, and we haven’t proven that we don’t need engineers.”
She believes that the focus should not be on eliminating technical talent, but on using AI to enhance the role of technical people. This can be achieved through abstraction, reducing tedious work, generating tests, writing documentation, and capturing the overall perspective of the developer experience.
Laycock said that in this fundamental shift, when you put anything or everything generated by AI into production, don’t forget to observe everything.
“Observe it. AI is integrating into all these areas, which is good, but we have to remember that it’s not deterministic. Before we finally find the solution everyone dreams of, more training and tuning might be needed for some more complex problems,” Laycock said.
Finally, she stated definitively: “I am sure people’s jobs will change, but as for how jobs will change and what changes are happening,” as an industry, we are far from figuring that out.
Author: Changzhang
Related reading:
Linus Torvalds Looks Back on 20 Years of Git
After the Crawler Crashed the Website, the Programmer Built a “Bomb” to Fight Back
Code War: Rust vs. C for Securing Billions of Devices!