my thoughts on Kiro
A few weeks ago, at the NYC Summit, AWS announced the launch of Kiro – a new IDE for AI-native and agentic development. What stood out most to me was not the software itself, but the workflow. AI is changing the way we think and how we work, and it would only be natural that the same would be true for the tools we use to code. Kiro is just at the start of transforming how we interact with software tools, and I want to share my thoughts.
I’ve been hearing a lot about vibe coding these days, from customers, coworkers, and friends alike (for those unfamiliar, vibe coding is a style of software development where AI agents generate the majority of the actual code, less hands-on-keyboard work for coders). As someone who does not come from a traditional software development background, Kiro is the perfect tool to add more structure to vibe coding, bringing up concerns or ideas I hadn’t considered, or helping me brainstorm how I might improve an idea before coding even begins.
Before I get into the actual workflow of the tool for those interested, I want to highlight a key feature of Kiro. A hook is an automation tool that performs a task based on an action done by the user. The simplest example of this is hitting save while coding (you can think of this similarly to saving a word document). You can create a hook based on ‘saving’ to automatically update documentation or generate a new unit test based on changes in the file. These tasks can be tedious for developers and devops teams to maintain and can slowdown deployment of production-ready code.
A perfect use case for Kiro came out of a conversation with, of all people, my father. He recently started working at a startup called Chainguard, a container security company. Chainguard provides secure container images that are free from common vulnerability and exposures (CVEs). Picture this: every time a developer saves a file, Kiro automatically checks the code against the provided Chainguard images for malicious code. This hook designed as a security guardrail is exactly where Kiro shines.
I do want to spend time exploring the details of Kiro’s workflow, but I’ll start by touching on how the development process of creating an AI agent works today. Most developers who craft agents send large, complex prompts to LLMs that process and output an agent workflow. This can be problematic for a lot of reasons, the main one being a lack of visibility into how the agent will be making decisions. Transparency and explainability are key concerns. LLMs make a lot of assumptions based on gaps in prompts and how it was trained, both of which can be hard to identify. To combat this, let’s walk through Kiro’s approach, what they are calling ‘spec (specification) driven development’.
Start by describing the agent decisions or workflow you are looking to solve. Where a tool like Cursor (or any other AI coding assistant) would begin generating code immediately, Kiro takes in your description and returns a set of requirements in natural language that it thinks will meet your description. You chat back and forth, giving feedback about the requirements you like, and where you want adjustments made.
Once you’re satisfied with the requirements, Kiro will begin to build a design doc, generating technical specifications, architecture diagrams, workflows and subtasks that will be used to achieve the agent logic you described. Again, this is an iterative process. Kiro will take feedback, clarifying any assumptions that it may have made incorrectly.
Finally, task creation begins. Kiro works in autopilot mode to create the code that will characterize your agentic workflow. This can be stopped at any time, and since you iterated beforehand with requirements and specifications, Kiro can adjust its output without having to start over from scratch.
Simple conceptually, but groundbreaking in terms of how a developer interacts with an AI coding assistant today. All the answers aren’t needed from the start - Kiro can help parse out the most important architecture requirements and how it will fit into a larger logical flow as you go. By deep diving into details, the agent coding experience is more enterprise-grade and production ready.
As I continue to work with my customers on their applications, I’m excited to see how Kiro can support their AI development process and meet them where they are. While Kiro remains in private beta, we are eager to get more feedback on the tool, make it faster, and support more LLM models.
—
Learn more about Kiro here