The rise of artificial intelligence has sparked a new era of agent-based software development, fundamentally changing how code is created, tested and deployed. 

While “vibe coding” — writing software through casual conversation with a coding agent — has captured the imagination of hobbyists, enterprise information technology requires a level of rigor and maintainability that standard chat windows often lack, according to Deepak Singh (pictured), vice president of developer agents and experiences at Amazon Web Services Inc.

“The challenge with [vibe coding] is five days after you’ve written an application, you’ve kind of forgotten why,” Singh told theCUBE. “In a team environment, it becomes even more interesting because three months later, nobody has any idea why you wrote the software you did.”

Singh spoke with John Furrier as part of theCUBE’s coverage of AWS re:Invent, for an exclusive interview on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed agent-based software development and Kiro, AWS’ bid to turn loose “vibe coding” chats into spec-driven, testable software that enterprises can actually ship.

Structuring agent-based software development

Since its announcement at the AWS Summit in July, Kiro — an intelligent assistant for software development — has seen rapid adoption, logging over 250,000 users in its first three months. However, the tool is evolving beyond simple code generation toward “spec-driven development,” an approach that models the behavior of Amazon’s own principal engineers, who focus on defining the problem before writing a single line of code, Singh explained.

“Senior engineers, our principal engineering community, [are] really good at helping people understand the problem, break it into smaller parts and maybe work with others to clarify exactly what they need to do,” he said.

Kiro allows developers to convert conversational intent into a formal specification, followed by a design and finally a set of executable tasks. This ensures that the convenience of AI-assisted coding does not come at the expense of documentation or structural integrity, Singh noted. Instead, this method goes beyond standard unit tests by using the agent to extract functional requirements from the specification and automatically generating tests to verify those properties.

“Here the agent is saying, ‘Here’s all the things that code should functionally do. We are going to actually test if it actually does that,’” Singh said. “[Meaning] your code is much more likely to be actually good and do what you want it to do.”

Developers can create specific profiles — such as an “AWS operator” or a “frontend developer” — that automatically load the correct tools, steering files and best practices for that specific domain, according to Singh. This flexibility allows teams to enforce security guidelines and architectural standards automatically, effectively shifting compliance to the earliest stages of development.

“You can switch back and forth between these custom agents that are specialized at other tasks and their manifestations of what you [would] like to do,” he explained. “It’s a different way of building. In some ways, it’s very familiar. In some ways, it’s unique.”

Ultimately, these tools are lowering the barrier to entry, allowing senior engineers to code more and bringing lapsed developers back into the fold, Singh noted. This means that more people can write more code — faster than ever. 

“I’ve seen senior engineers who have written more code in the last six months than they had written in the three years [previously],” he said. “These tools allow them to express themselves in ways that they just couldn’t before.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of AWS re:Invent: 

Photo: SiliconANGLE