Thoughtworks CTO: AI accelerates software debt, not solves it

AI is accelerating software debt, not solving it

AI is acting as a “funhouse mirror” for software development, amplifying both good practices and existing problems. That’s the warning from Thoughtworks CTO Rachel Laycock, who argues that without strong engineering fundamentals, AI’s ability to generate code quickly just becomes a debt accelerator.

“Sure, you can get an agent or a swarm of agents to build lots of software for you, but is that software secure? Is it optimized for cost?” Laycock asked in an interview with The New Stack. She worries the industry is overlooking these cross-cutting concerns in the rush to adopt AI.

Velocity is not the real bottleneck

The core issue is that writing code was never the primary bottleneck in software delivery. The 2025 DORA report positions AI primarily as an amplifier. If teams lack solid DevOps, testing, and security practices, AI simply makes things worse, faster.

“Building an app is easy. It’s always been easy if you’re a coder. Now it’s easy if you’re not a coder,” Laycock said. “But an app is not distributed, complex business software.”

She points to the enduring importance of fundamentals that distinguish high-performing teams:

• DevOps and CI/CD best practices
• Security and testing guardrails
• Observability and progressive delivery
• Modular, well-architected systems

The influx of new builders requires stewardship

AI is lowering the barrier to entry, allowing more people to create software. Laycock sees this as positive but warns that the industry’s response will be critical. A protectionist mindset that uses AI to replace junior engineers is a mistake.

“Instead of being like: ‘Hey, they don’t know what they’re doing, stop it,’ we should think about: How can we become good stewards?” she said. Senior technologists need to explain the “why” behind engineering rigor to newcomers who haven’t learned the hard lessons of distributed systems.

This requires empathy, especially for subject matter experts from outside tech who can bring valuable user perspective but lack technical background. “We don’t talk about the why. We just have this assumption: Of course, you have to care about that thing because we’ve learned that,” Laycock warned.

Programming languages may diverge for agents

The rise of AI agents is forcing a reconsideration of programming languages themselves. Code must now be optimized for both humans and AI, which could lead to a split in how software is built.

“We might see new types of languages, maybe very heavily typed languages, that lean more into having a structured approach for programming,” Laycock said, referencing discussions at a recent Thoughtworks retreat. One key question emerged: “How much fiddling with the code are we actually going to do in the future?”

As the need for humans to read code diminishes, languages could become more efficient for AI agents to parse than for people. This leads to a potential divergence:

• Some teams prioritizing human-readable code.
• Others building software in new ways optimized solely for agents.

Not all software should be built the same way

The industry is exploring new development paradigms, like “moldable development” or the principles in the Outcome Engineering Manifesto (O16g), which prioritize business results over software delivery. Thoughtworks itself has launched AI/works, an agentic platform using AI and domain-driven design.

Laycock argues for a risk-based approach. Enterprises will need to decide where to allow more AI generation versus where to keep humans firmly in the loop.

“There’ll be other software, which is extremely high business value, high-impact, customer facing, where we want more humans in there, where we want the software to be more human-readable, more modularized, well-understood, strong guardrails,” she said. The consensus from the retreat was not that AI will replace humans soon, but that its application must be guided by the value and risk of the software being built.