Amazon Web Services (AWS) has added two additional capabilities to its Kiro artificial intelligence (AI) coding tool that promise to make it simpler to both create higher quality software in the first place and fix any bugs that might have inadvertently been included.
At the core of that capability is a specifications-based approach to enabling AI agents to perform tasks that AWS has embedded within Kiro. That capability is then used to provide developers with access to Kiro Powers, a suite of AI agents that have been trained to automate a task, such as reviewing code, within the context of a narrow set of guidelines that prevent the AI agent from performing tasks beyond a specific purview.
AWS has now added a Design-first specification to Kiro that ensures code is generated within a set of best practices defined by AWS and a Bug Fix specification that enables an AI agent to automatically resolve issues in code long before software is deployed in a production environment.
Amit Patel, director of software development for Kiro at AWS, said the Bug Fix specification essentially extends those principles further by making it possible for developers to assign an AI agent to automatically resolve any issues that arise as code is being developed. Ideally, developers should define the detailed scenario that caused the error, not just the error message itself, said Patel. Just as importantly, they will need to explicitly state what should not be modified to ensure other bugs are not introduced as code is modified, he added.
Kiro will also generate property-based tests (PBTs) to first verify there is a bug. It then generates additional tests to verify a bug has not only been fixed but also to determine that the behavior of the application has not somehow been changed.
In general, most application developers are going to start with Design-first specification unless the behaviors and outcomes of the application that needs to be built are very well understood, noted Patel. Developers are presented with a high-level design for system diagrams and components or, alternatively, can invoke low-level design for algorithms and function signatures.
Kiro then derives the requirements from the design document, which is then used to assess the feasibility of the proposed project.
Hopefully, the number of bugs generated during that process will be minimal but the overall goal is to make it simpler to generate higher quality code with fewer bugs that pass muster with the software engineers that typically review it before being added to a production environment, noted Patel.
It’s not clear how pervasively DevOps teams are employing AI coding tools, but a recent Futurum Group survey finds a full 60% of respondents said their organization is now actively using AI to build and deploy software. The top areas of investment over the same period are AI Copilot/AI code tools (38%), AI agent development (37%), AI-assisted testing (37%), followed closely by DevOps (37%), automated deployment (34%), software security testing (31%).
Mitch Ashley, vice president and practice lead for software lifecycle engineering at the Futurum Group, said advances such as the Design-first and Bug Fix specifications suggest AI coding tools are rapidly maturing to meet developers where most enterprise work happens: existing codebases and surgical bug fixes. This reflects competitive pressure to prove agent tooling works at production reality.
Additionally, the ability to specify “Unchanged Behavior” is operationally significant because it requires developers to define what must not change to directly address the regression risks that make engineering teams distrust autonomous agents with production code, he added.
Ultimately, it’s soon not going to be a question of whether to use AI coding tools but rather the degree of faith that application development teams will have in them.