Amazon Web Services has unveiled the Kiro autonomous agent, a new class of AI system that can independently handle software development tasks for hours or days at a time — no constant human guidance required. Announced by AWS CEO Matt Garman during his keynote at AWS re:Invent 2025 in Las Vegas on December 2, Kiro marks a significant leap beyond AI coding assistants that respond to prompts and wait for the next instruction.
The agent is part of a broader AWS initiative called frontier agents — a new category of AI systems defined by three characteristics: they are autonomous, massively scalable, and capable of working independently over extended periods. Alongside Kiro, AWS simultaneously introduced the AWS Security Agent and the AWS DevOps Agent, each targeting a specific phase of the software development lifecycle.
Kiro is not a chatbot. It does not respond to a single prompt and stop. Assign it a task from the backlog, and it independently plans the implementation, writes code across multiple repositories, runs tests, and opens pull requests — all while the developer works on something else entirely.
What Makes Kiro a Frontier Agent
The term “frontier agent” carries a specific technical meaning in AWS’s framework. These are not AI tools that assist with individual steps. They are systems that take a high-level goal, figure out how to achieve it, execute the work across whatever tools and repositories are needed, and keep going — sometimes for multiple days — without asking for help at every turn.
Three properties define a frontier agent in AWS’s classification. Autonomy means the agent determines its own path to a goal once that goal is stated. Scalability means it can run multiple concurrent tasks and distribute work across specialized sub-agents. Independent operation means it can maintain context and progress over long stretches of time without losing its place or requiring a human to re-explain the context.
Kiro’s persistent context capability is the technical foundation that makes the rest possible. Unlike AI coding tools that reset after each conversation, Kiro maintains awareness across all repositories, sessions, tickets, and pull requests it has touched. Every code review comment, every architectural decision, every standard the team follows — Kiro absorbs all of it over time and applies it going forward without being reminded.
How Kiro Handles Real Development Work
Garman used a concrete example during the re:Invent keynote to illustrate what Kiro is capable of in practice. A developer needs to update a critical piece of code that is referenced by 15 separate internal systems. Instead of assigning and verifying each update one at a time, the developer assigns Kiro the full task in a single prompt. Kiro identifies all 15 dependencies, plans and executes the coordinated updates, and opens individual pull requests for each — all without manual intervention between steps.
Tasks can be assigned to Kiro in two ways. Developers can create tasks directly from the Kiro web interface at kiro.dev, or they can trigger the agent through GitHub Issues by writing /kiro in a comment. The second method is particularly significant because it fits directly into how development teams already work, requiring no change to existing workflows.
Every task Kiro executes runs inside an isolated sandbox environment. The agent opens pull requests for human review and never merges changes automatically. AWS has also built in comprehensive logging so developers can audit every action the agent took. Branch protection is recommended for main and sensitive branches to prevent direct pushes while the agent is executing tasks. For teams exploring human-in-the-loop AI governance frameworks, Kiro’s enforced review checkpoints represent a practical implementation of that model in a live development environment.
Team Mode: Kiro as a Shared Development Resource
Individual developers are not the only intended users. AWS designed Kiro to function as a shared team resource that builds a collective understanding of the entire codebase, the product roadmap, and the standards the team enforces. It connects directly to Jira, GitHub, and Slack, pulling context from wherever the team’s work actually lives.
When a developer leaves feedback on a pull request — something like “always use our standard error handling pattern” — Kiro retains that instruction and applies it to all future work automatically. Over time, the agent’s understanding of the team’s preferences deepens with every ticket closed, every review comment submitted, and every architectural decision logged. The result is an AI that gets more useful the longer a team works with it, not one that resets to zero with every session.
This model is a fundamental departure from how most AI development tools operate. NVISIONx, an early access partner, reported that Kiro’s ability to maintain deep contextual awareness across sessions and handle asynchronous project workloads showed clear potential to accelerate product teams working across multiple repositories simultaneously.
The Other Two Frontier Agents: Security and DevOps
Kiro operates alongside two companion frontier agents, each addressing a different part of the software development lifecycle. The AWS Security Agent acts as a virtual security engineer, performing threat modeling, secure architecture reviews, static application security testing, dependency checks, and penetration testing-style analysis on demand. It can run continuously across an environment, scanning for configuration drift and automatically opening tickets when issues surface.
The AWS DevOps Agent targets operational reliability. Western Governors University, which serves approximately 200,000 students on a 24/7 platform, reported that after integrating the DevOps Agent with Dynatrace, the system could autonomously investigate the full technology stack when performance issues occurred — identifying root causes from third-party API failures, networking problems, or application-level errors without waiting for a human to begin the investigation.
All three agents are built on Amazon Bedrock AgentCore, the infrastructure layer that provides memory, policy enforcement, evaluations, and observability controls. Before executing potentially dangerous actions, AgentCore runs policy checks and automated reasoning to verify the action is permitted — a safeguard that AWS is positioning as central to enterprise-grade trust in autonomous systems. Teams working on AI governance and compliance will find the AgentCore policy layer directly relevant to their regulatory obligations.
Spec-Driven Development and How Kiro Learns Your Codebase
The technical foundation beneath Kiro’s autonomy is a concept AWS calls spec-driven development. As the agent codes, it asks developers to confirm or correct its assumptions, building up a specification layer that documents how the team expects things to work. Future tasks are executed against those specifications rather than from scratch, which is why Kiro’s output becomes more aligned with team standards the longer it is in use.
The agent learns from scanning existing code, reading documentation, processing pull request histories, and absorbing review feedback. This is how it develops what AWS describes as a “collective understanding” of the codebase — not through a one-time onboarding process, but through continuous accumulation of context across every interaction. Developers building on top of cloud infrastructure or evaluating AI-driven platforms from an investment perspective will recognize this self-improving loop as the core differentiator that separates frontier agents from conventional AI coding tools.
Availability, Pricing, and Access
Kiro autonomous agent is currently in preview, rolling out gradually to Kiro Pro, Pro+, and Power subscribers. Team access requires an invitation through a separate waitlist. There is no additional cost during the preview period, though usage is subject to weekly limits that reset each week. AWS has stated it will communicate pricing before the preview period ends.
To activate the agent, users connect their GitHub account through the Kiro web interface and configure repository access. AWS explicitly advises selecting only trusted repositories and protecting sensitive branches. The agent only functions in repositories where the Kiro Agent GitHub App is installed and the user’s GitHub account holds the appropriate access rights.
Developers who prefer command-line workflows can also access Kiro through the CLI, which supports agent capabilities alongside its standard terminal-based development features. The AWS SDK also allows programmatic invocation, making it possible to embed frontier agents into existing internal tooling and automation pipelines.
AWS CEO’s Vision: AI That Works Like a Team Member
Garman was direct about what AWS is trying to accomplish. “You simply assign a complex task from the backlog and it independently figures out how to get that work done,” he said during the re:Invent keynote. “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.”
AWS acknowledged in its own communications that earlier AI coding tools introduced friction rather than reducing it. Developers found themselves acting as the connective thread holding AI-assisted work together — re-contextualizing tasks, manually coordinating cross-repository changes, and collating information across tickets and pull requests that the tools could not track on their own. Garman admitted that even with the earlier Kiro IDE, efficiency gains were “more incremental than transformative” during the first weeks of adoption. The frontier agent layer, in his framing, is the final component that resolves that friction.
The launch of Kiro autonomous agent, AWS Security Agent, and AWS DevOps Agent at re:Invent 2025 signals a concrete shift in how AWS is positioning AI in the enterprise software development cycle — not as a tool that developers use, but as a class of independent contributors that work alongside them. Whether that promise holds at scale and under production conditions will become clear as the preview expands and more teams gain access over the coming months. What is already certain is that the definition of what an AI coding system can do has moved significantly further than where it stood twelve months ago.
