Amazon Unveils Three AI Agents, Including 'Kiro', Capable of Autonomous Coding for Days
Amazon Web Services (AWS) has unveiled three cutting-edge AI agents, dubbed 'frontier agents', during its AWS re:Invent event on Tuesday. One of these agents, named 'Kiro', is designed to learn and adapt to an individual's work style, enabling it to operate autonomously for extended periods, potentially spanning days.
These agents are tasked with various responsibilities, such as writing code, conducting security audits, and automating DevOps processes to prevent issues during code deployment. The preview versions of these agents are now accessible for testing.
The most intriguing aspect of AWS's announcement is the claim that the 'Kiro autonomous agent' can function independently for days at a time. Kiro is an AI coding assistant built upon AWS's existing Kiro tool, which was introduced in July. While the original Kiro was primarily used for rapid prototyping, the new Kiro is designed to produce fully functional, operational code that can be deployed.
To achieve this, Kiro employs a technique called 'spec-driven development'. As Kiro codes, it requires human input to instruct, confirm, or correct its assumptions, thereby creating specifications. The autonomous agent then observes how the team operates using various tools, including code analysis, to learn and adapt.
AWS CEO Matt Garman highlighted the capabilities of Kiro during his keynote presentation. He stated that Kiro can be assigned complex tasks from a backlog and will independently determine the best approach to complete the work.
"Kiro learns how you prefer to work and continually enhances its understanding of your code, products, and team standards over time," Garman explained.
One of the key advantages of Kiro is its ability to maintain context across sessions. Unlike other AI models that may run out of memory or forget their assigned tasks, Kiro can handle tasks and work independently for hours or even days with minimal human intervention.
Garman provided an example of Kiro's capabilities: it can be tasked with updating critical code segments used across 15 different corporate software applications. Instead of assigning and verifying each update individually, Kiro can be instructed to fix all 15 updates in a single prompt.
To complete the automation of coding tasks, AWS has also developed the AWS Security Agent and the DevOps Agent. The Security Agent identifies security issues during code development, tests them, and suggests fixes. The DevOps Agent, on the other hand, automatically tests new code for performance and compatibility issues.
While Amazon's agents are not the first to claim long work periods, it's worth noting that OpenAI's GPT-5.1-Codex-Max model is also designed for extended runs, up to 24 hours. However, developers still face challenges with context windows, as LLMs may produce inaccurate or hallucinated results, requiring constant supervision.
Despite these challenges, the expansion of context windows is a significant step towards making AI agents more like co-workers. Amazon's technology represents a substantial advancement in this direction.
