AWS Kiro tackles AI agent reliability with math-based bug detection before code is written

AWS targets AI slop with new spec check in Kiro coding tool, amid scrutiny of agent reliability

Amazon Web Services is adding a feature to its Kiro AI coding tool designed to mathematically prove that software requirements are free of contradictions and gaps before any code gets written, addressing one of the core risks of AI-assisted software development. The feature, called Requirements Analysis, is designed to catch the kind of bugs that can often be the hardest to spot and most expensive to fix -- problems that start not in the resulting code but in the initial requirements that define what the software is supposed to do. The announcement Tuesday morning comes three months after Amazon publicly pushed back on a Financial Times report that its AI coding tools contributed to AWS outages, an episode that highlighted the risks of giving AI agents too much autonomy in software development. It also comes a day after AWS hired former Microsoft exec Shawn Bice to return to Amazon as VP of AI Services leading its Automated Reasoning Group, the team behind the new feature. Bice will report to Swami Sivasubramanian, Amazon's VP of Agentic AI. Requirements Analysis combines large language models with an automated reasoning engine called an SMT solver. The LLM translates natural-language requirements into formal logic. The solver then checks those requirements by mathematically proving whether they contradict each other or leave gaps that could be filled in erroneously by the AI coding tool -- a common problem as AI increasingly generates software faster than developers can review it. "Every vague prompt produces a vague spec or plan, and the AI agent implementing that spec produces code full of undisclosed decisions made on your behalf, without your awareness or agreement," wrote AWS applied scientists in a blog post accompanying the news. Kiro competes in a crowded and fast-growing market for AI coding tools that includes Cursor, GitHub Copilot, Anthropic's Claude Code, Google's Antigravity, and OpenAI's Codex. While those tools have increasingly added planning and agent workflows alongside code generation, Kiro has built its identity around a spec-first approach that requires developers to formalize their intent before the AI starts building. AWS also announced two other Kiro features designed to speed up the development process. * Parallel Task Execution runs independent coding tasks concurrently rather than sequentially, cutting implementation times for large projects by roughly 75 percent, according to the company. * AWS says a new Quick Plan mode lets developers skip the step-by-step approval process for well-understood features, generating a full set of requirements, design, and tasks in one pass.

AWS Kiro accelerates software development by proving code correctness before it gets to work - SiliconANGLE

AWS Kiro accelerates software development by proving code correctness before it gets to work Amazon Web Services Inc. is trying to get rid of the bottleneck between architectural planning and code execution with a number of upgrades to its artificial intelligence software development tool Kiro. The upgrades, which are all rolling out today, include Parallel Task Execution and streamlined Quick Plan workflow capabilities designed to help developers move faster. They're joined by a new Requirements Analysis engine designed to catch issues with the code before a single line is written. In a blog post, AWS Product Manager Ankit Sharma and Principal Engineer Richard Threlkeld explained that Kiro is focused on "spec-driven development" that's designed to deliver higher-quality code implementations. However, it's a cautious approach that sacrifices something that many organizations prioritize - developer velocity. For instance, if Kiro is fed a feature specification with 10 tasks and six of them are all independent of one another, with different endpoints, files and no shared state, it will complete them sequentially, one after another, rather than doing them all at once. Moreover, for projects where the user already knows the scope and constraints, Kiro's step-by-step approval flow is probably overkill. But on the flip side, there are cases where a deceptively simple feature prompt "may include many unstated assumptions and ambiguities that can take the implementation in the wrong direction." Today's updates represent an effort by AWS to streamline Kiro's development process while making sure that implementations never end up going in the wrong direction. The new Requirements Analysis engine employs a three-stage neurosymbolic pipeline that begins with large language models rewriting the user's vague requirements into testable criteria, which can then be translated into formal logic. That logic is then submitted to a Satisfiability Modulo Theories solver - essentially, an automated reasoning engine. Unlike standard LLMs, which work by predicting the next word in a sequence, the SMT solver uses mathematics to prove if contradictions exist. If it discovers that two requirements are logically incompatible - for instance, if a rule on page one mandates a hard delete and a rule on page 10 implies a soft delete - the solver will identify that conflict as a mathematical impossibility. Kiro will then surface these findings in plain language that developers can understand, so they can quickly come up with a fix. As for the Parallel Task Execution, it's meant to solve the problem outlined above when Kiro is fed a specification with multiple independent tasks. It works by analyzing the dependency graph of each new project, identifying which tasks do not share a state, endpoints and files. These will then be run concurrently in isolated contexts, speeding up the overall development time for large specifications from more than an hour to as little as 15 minutes, AWS said. To complement this, AWS has developed Quick Plan, which is a kind of fast-track mode for building well-understood features. Instead of asking for approvals step-by-step, Kiro will simply ask a bunch of clarifying questions up front, before heading off to generate the entire stack in one go. Today's updates could have significant implications for autonomous AI agents. By applying the same mathematical rigor used in hardware design to software engineering, Kiro should be able to eliminate many of the "hallucinations" that continue to plague AI development. They should also help to make coding agents feel a bit smarter. In many ways, existing AI coding bots feel as if they lack any common sense. If a developer feeds it a blueprint that shows a staircase leading towards a solid brick wall, it won't question that design, but just build it exactly as it's shown. With the Requirement Analysis engine now present, coding bots can act more like structural engineers who will notice the problem before it even gets started on the foundations.