The launch of Kiro, a development environment offered by Amazon Web Services Inc., provides a new path for software engineers to turn ideas into production-ready code, moving beyond “vibe coding” and into a specification-driven model.

Caylent’s Randall Hunt spoke with theCUBE about the development environment during the AI Cloud Leaders event.

Kiro’s combination of “spec coding” with integrated AI agents is designed to enable developers to focus more closely on systems design. This will give Kiro an advantage over other integrated development environments, according to Randall Hunt (pictured), chief technology officer of Caylent Inc., an AWS Premier Partner that helps organizations build, scale and optimize cloud solutions through agile co-delivery and deep cloud expertise.

“One of the things that sets it apart from other IDEs like Cursor or Windsurf for Codeium … is a spec-driven development model,” Hunt said. “The spec-driven mode goes through a set of requirements, collecting requirements, you edit them, you click through them, then it goes into design. Something that I’ve been saying over the last few days is the people who are going to be successful in the world of LLMs are the ones who have agency. And I believe that Kiro is the agent for people with agency.”

Hunt spoke with theCUBE’s John Furrier at the theCUBE + NYSE Wired: AI + Cloud Leaders Media Week event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed evolving trends in AI deployment including agents, models and systematic evaluation.

Agents support development environment

Kiro’s integration with AI agents to perform spec-driven coding tasks highlights the growing role of autonomous software in the enterprise. Kiro provides an agentic chat function for coding tasks within a file and agents can be connected to external open-source tools.

The need for periodic human oversight still remains, although agents can operate now for longer periods, according to Hunt.

“Right now, the best frontier models every 10 minutes require human intervention,” Hunt said. “But that 10-minute frontier keeps pushing out. The amount of time that an agent can autonomously execute keeps getting larger and larger. So, we’re probably less than a year away from an agent that can autonomously execute by itself without error for 24 hours.”

The introduction of a new AI-based development environment by AWS also underscores the increased reliance on multiple models for accomplishing enterprise tasks. One report has found 3,000% growth in enterprise AI and machine learning usage, and the Hugging Face Hub now hosts over 1.7 million AI models in its repository.

“I see companies betting on one model, and I think that is such a terrible approach. I think you need to build a platform,” Hunt said. “The models are moving so quickly. One model family could develop capabilities that immediately allow you to differentiate yourself from your competitors. It’s better to build a platform that can constantly and consistently take advantage of the state of the art.”

Companies are also finding that AI usage requires a process that can systematically evaluate the performance of large language models and their applications. Hunt cited a recent project with a customer who wanted to build video semantic search from millions of hours of footage where evals became a key element.

“We fed this into the Amazon Nova models and we were able to process these petabytes of image data using Bedrock batch,” Hunt explained. “Your eval set becomes every different kind of video. When we enriched that prompt and used our eval set to really measure how accurate the description and the understanding of the video was, we got much, much better results. It’s really kind of like a continuous integration/continuous deployment system for LLMs.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of theCUBE + NYSE Wired: AI + Cloud Leaders Media Week event:

Photo: SiliconANGLE