Amazon Web Services announced three new AI agents it calls “frontier agents” on Tuesday. Preview versions of these agents are available now. Each agent handles a different set of tasks, such as writing code, managing security processes like code reviews, and automating DevOps tasks to prevent incidents when pushing new code live.
Perhaps the most significant claim from AWS is about the frontier agent named “Kiro autonomous agent,” which it promises can work independently for days at a time. Kiro is a software coding agent built upon AWS’s existing AI coding tool, also called Kiro, which was announced in July. While the earlier tool was useful for prototyping, it was designed to produce operational code ready to be deployed. To create reliable software, the AI must adhere to a company’s coding specifications. Kiro achieves this through a method called “spec-driven development.”
As Kiro writes code, it requires a human to instruct, confirm, or correct its assumptions, thereby helping to create those specifications. The Kiro autonomous agent learns how a team works by scanning existing code and through other training methods. AWS states that after this learning phase, it can operate on its own.
AWS CEO Matt Garman introduced the product during his keynote at AWS re:Invent. He explained that you can assign a complex task from the backlog and the agent will independently figure out how to complete it. He added that it learns your preferred working style and deepens its understanding of your code, products, and team standards over time.
Amazon says Kiro maintains “persistent context across sessions,” meaning it does not run out of memory or forget its assigned tasks. Therefore, it can be given a task and work autonomously for hours or days with minimal human intervention. Garman illustrated this with an example where Kiro could update a piece of critical code used by 15 different corporate software applications, fixing all of them from a single prompt instead of requiring separate, verified updates.
To complement the automation of coding tasks, AWS developed the AWS Security Agent. This agent works independently to identify security issues as code is written, tests the code afterward, and suggests fixes. The third agent, the DevOps Agent, automatically tests new code for performance problems or compatibility issues with other software, hardware, or cloud configurations.
It is important to note that Amazon’s agents are not the first to promise extended operational periods. For instance, OpenAI stated last month that its agentic coding model, GPT-5.1-Codex-Max, is also designed for long runs of up to 24 hours.
Furthermore, it is not entirely clear that the primary challenge for agent adoption is the context window, or the ability to work continuously. Large language models still struggle with hallucinations and accuracy problems, which developers say turns them into “babysitters” who must frequently verify the AI’s work. As a result, developers often prefer to assign short tasks and check results quickly before proceeding.
However, for AI agents to become true co-workers, their context windows must expand. Amazon’s technology represents another significant step in that direction.