AWS Is Trying to Make Kiro the Most Trustworthy AI Coding Tool in the Room
The AWS logo displayed at the company's 2024 re:Invent conference on Dec. 3, 2024. Photo credit: Ken Yeung

In software development, the costliest mistake rarely lives in the code. It lives in the requirements that came before it. Typically, the issue arose because a developer interpreted the specs differently from the author—and that’s human-to-human. Imagine how much more challenging this can be when vibe-coding. That’s why Amazon Web Services (AWS) is updating its Kiro integrated development environment (IDE) to address this issue and ensure that the code generated is trustworthy.

“Addressing these bugs in your requirements is so important because the errors or incompleteness gaps, sort of logical inconsistencies, just multiply as they get missed throughout the software development life cycle,” Mike Miller, AWS’s director of AI product management, said to The AI Economy in an interview. “What we want to try to do is address…one of the most expensive classes of bugs at the cheapest possible moment.”

Trust Through Neurosymbolic AI

What AWS is doing goes well beyond helping developers write better spec documents. It’s about mathematically verifying that what they’re asking AI to build is actually buildable—before a single line of code is generated. The approach relies on Neurosymbolic AI, a technique that combines the language fluency of large language models with the provable certainty of formal mathematical logic.

The concept isn’t new—its roots trace back to the earliest days of AI research. But AWS has quietly been applying it for the past decade under what it calls automated reasoning. Miller explained that the company has used it internally to verify cryptographic algorithms, validate access control policies, and mathematically guarantee that an S3 bucket marked as private can never be reached from the public internet.

The integration with Kiro marks the first time AWS has brought that same rigor directly to developers as a hands-on, consumer-facing tool.

“We’re really excited about this Neurosymbolic technology because we see applications of it to a number of different capabilities, both in Kiro and across other user-facing tools at AWS,” Miller said. “We really see this as a key capability to help our customers achieve more trustworthy AI.”

How It Works With Kiro

Software engineers often treat spec documents, no matter whether they’re written by hand or using an LLM, as a first draft. But what happens when an LLM comes up against something ambiguously written? Vague or contradictory specs don’t just slow things down; they get implemented, and it becomes someone else’s problem down the road.

Neurosymbolic AI becomes visible in Kiro’s new Requirements Analysis feature. First, the LLM reviews the acceptance criteria and rewrites anything too vague to be testable—removing ambiguous language and tightening the level of detail. Next, those refined requirements are translated into a formal mathematical representation. Kiro samples multiple translations of the same requirement and looks for divergence. If the translations cluster around a single consistent interpretation, the requirement is unambiguous. However, if they scatter, that’s a flag—the LLM was guessing at the meaning, which means a human would too. Kiro identifies exactly where the interpretations split and serves up a plain-language question: here’s the ambiguity, which did you mean: A or B?

In the final phase, an automated reasoning engine analyzes the full set of requirements together, checking for contradictions between rules that looked fine in isolation, finding gaps where certain situations have no defined behavior, and flagging what Miller called “vacuous requirements”—rules that don’t actually constrain anything and would only produce unnecessary code.

AWS says roughly 60 percent of draft requirements across 35 internal Kiro projects required refinement before they were ready to generate valid code.

Credit: AWS
Credit: AWS

Other Kiro Updates

Requirements Analysis isn’t the only thing changing in Kiro. AWS is also shipping two updates aimed at speed:

Run Tasks in Parallel

With “Run all Tasks,” Kiro can run multiple tasks simultaneously rather than sequentially. When you kick off a spec, Kiro maps out which tasks depend on each other and which ones don’t—and runs the independent ones concurrently. Each task operates in its own isolated context, so there’s no interference between them—if one fails, the others keep going. AWS reports implementation times for large specs have decreased by 75 percent from over an hour to about 15 minutes in some cases.

Quick Plan for Specs

A specification workflow typically has three phases: requirements, design, and tasks—each needing approval before moving on to the next stage. This is useful when working through something unfamiliar. But when it comes to building well-documented features, the process can be laborious.

To move things along, AWS has introduced “Quick Plan” in Kiro. Developers are first asked clarifying questions about their app’s scope and constraints. Then the AI will generate all three phases in a single pass, resulting in a task list ready to be built. The underlying documentation is still generated and saved, but developers aren’t required to stop and approve each piece before the next phase begins.

The Investment in Being Correct

Today’s updates arrive alongside a leadership change inside AWS’s Automated Reasoning Group (ARG). Shawn Bice, Splunk’s former president of products and technology, has been tapped to lead the company’s investment in Neurosymbolic AI. AWS’s Vice President of Agentic AI, Swami Sivasubramanian, described him in an internal employee memo as someone who brings “decades of experience building and operating cloud services at massive scale, deep customer obsession, and a track record of attracting and developing world-class talent.”

His predecessor, Scott Wiltamuth, who has made significant contributions to ARG, will remain on the team focused on what AWS describes as its “highest-leverage technical problems.” The transition signals that AWS is treating automated reasoning not as a research curiosity but as a product priority.

When asked about future plans for Neurosymbolic AI beyond Kiro, Miller said he couldn’t speak about specific features but was direct about the direction. Requirements Analysis, he said, is “one step in our continued investment” in the space—one he expects to expand as agentic AI becomes more embedded in everyday development workflows.

For developers weighing what that means in practice, Miller’s message is straightforward: speed alone is no longer enough. “As AI writes more code, the human role should be shifting upstream,” he said. “The highest value work is going to come from defining what to build with precision, and not necessarily getting stuck on how to build it.”

Updated on 5/12/2026: Corrected Scott Wiltamuth’s role at ARG—he did not found the ground, that was Byron Cook, AWS’s vice president and distinguished scientist. However, Wiltamuth has played an oversized role in scaling the team.