Amazon Web Services (AWS) unveiled three new AI agents on Tuesday, collectively branded as "Frontier Agents," including one specifically designed to learn your work patterns and operate autonomously for multiple days.
Each agent is tailored to distinct tasks—such as writing code, handling security workflows like code reviews, and automating DevOps operations to prevent incidents during new code deployments. Preview versions of these agents are now available. Among AWS’s most notable claims is that its “Kiro Autonomous Agent” can function independently for several days.
Kiro builds upon AWS’s existing AI coding tool, also named Kiro, which was launched in July. While the original tool supports ambient coding—essentially rapid prototyping—its goal is to produce production-ready code or deploy software reliably. To ensure code quality, the AI adheres to a company’s specific software development standards through an approach AWS calls “specification-driven development.”
As Kiro codes, it engages human users to guide, confirm, or correct its assumptions, thereby co-creating those specifications. The Kiro Autonomous Agent further refines its understanding by observing how teams work across various tools—analyzing existing codebases and workflows. According to AWS, this enables it to operate independently thereafter.
“You simply assign it a complex task from your backlog, and it figures out how to complete the job on its own,” AWS CEO Matt Garman promised during the keynote address at AWS re:Invent on Tuesday while introducing the new offering.
“It actually learns how you prefer to work and deepens its understanding over time—of your codebase, your product, and the standards your team follows,” he added.
Amazon emphasizes that Kiro maintains “persistent context across sessions,” meaning it doesn’t lose memory or forget its objectives. As a result, the company asserts that Kiro can be assigned tasks and work autonomously for hours or even days with minimal human oversight.
Garman illustrated this capability with a scenario involving the update of a critical code segment used across 15 internal applications. Instead of requiring developers to manually assign and verify each individual update, Kiro could address all 15 in a single prompt.
To further automate secure coding, AWS also introduced the AWS Security Agent—an independent agent that identifies security vulnerabilities during code creation, performs post-write testing, and suggests remediations. Completing the trio is the DevOps Agent, which automatically evaluates new code for performance bottlenecks or compatibility issues with other software, hardware, or cloud configurations.
Of course, Amazon isn’t the first to claim extended autonomous operation for AI agents. For instance, OpenAI recently stated that its agent-based coding model, GPT‑5.1-Codex-Max, is engineered for continuous runs of up to 24 hours.
However, it remains unclear whether the primary bottleneck for such agents lies in context window limitations—the ability to work continuously without interruption. Large language models still grapple with hallucinations and accuracy issues, often forcing developers into a supervisory “babysitting” role. Consequently, many prefer assigning short, verifiable tasks before proceeding further.
Nonetheless, significantly expanding context windows is essential before AI agents can truly function like human teammates. Amazon’s latest advancements represent another substantial step toward that future.
