Amazon previews 3 AI agents, including ‘Kiro’ that can code on its own for days

Amazon Web Services on Tuesday announced three new AI agents, which it refers to as “frontier agents,” including one designed to learn a user’s workflow preferences and operate independently for days at a time.

Each agent handles different responsibilities, such as writing code, conducting security processes like code reviews, and automating DevOps tasks to prevent incidents when new code is deployed. Preview versions of all three agents are available starting today.

The most notable and boldest claim from AWS concerns the frontier agent called the “Kiro autonomous agent,” which it says can operate independently for extended periods, even days.
Kiro is a software agent built on top of AWS’s Kiro AI coding tool, announced in July. While the earlier version supported vibe coding (essentially rapid prototyping), it was built to produce production-grade, operational code. To deliver reliable code, the AI must follow a company’s coding standards. Kiro accomplishes this through a method called “spec-driven development.”

As Kiro writes code, humans guide the process by instructing, confirming, or correcting its assumptions, which, in turn, generate specifications. The autonomous Kiro agent learns further by observing how teams work within their tools and by scanning existing codebases. AWS says that once trained, it can operate independently.

“You simply assign a complex task from the backlog, and it independently figures out how to get that work done,” AWS CEO Matt Garman said while introducing the product during his keynote at AWS re: Invent on Tuesday.

“It actually learns how you like to work, and it continues to deepen its understanding of your code and your products and the standards that your team follows over time,” he added.

According to Amazon, Kiro maintains “persistent context across sessions.” In practical terms, this means it does not lose memory or forget tasks. It can therefore be assigned work and continue executing it for hours or days with minimal human oversight.

Garman shared an example involving a piece of critical code used by 15 different corporate software systems. Instead of assigning and reviewing each update individually, Kiro could be instructed to update all 15 at once.

To support full automation of coding workflows, AWS also revealed the AWS Security Agent. This agent operates independently to identify security vulnerabilities in real time as code is written, test the software, and suggest fixes. Completing the trio is the DevOps Agent, which automatically evaluates new code for performance issues, compatibility with other components, and alignment with cloud configuration.

It’s worth noting that Amazon isn’t the first to promise long-running agentic capabilities. For instance, OpenAI said last month that GPT-5.1-Codex-Max, its agentic coding model, can operate continuously for up to 24 hours.

However, context window limitations may not be the only bottleneck to widespread agent adoption. LLMs still face issues such as hallucinations and inaccuracies, forcing developers to act as “babysitters.” As a result, many prefer assigning smaller tasks that they can quickly verify.

Still, expanding context windows is essential before agents can function more like human co-workers. Amazon’s advancements represent a significant step toward that goal.