AWS brought its full agentic AI product stack to the Jacob K. Javits Convention Center on Wednesday for AWS Summit New York City 2026, a sold-out, free one-day conference where the central argument of the 90-minute keynote was direct: the era of AI tools that respond to prompts is giving way to AI agents that act — and the infrastructure required to run those agents safely in production has, until now, been missing. The keynote, led by Dr. Swami Sivasubramanian, Vice President of Agentic AI at AWS, and Chet Kapoor, Vice President of Security Services and Observability, ran from 11:00 AM to 12:30 PM ET and was available globally via free livestream. For developers who could not secure a badge — in-person registration closed at maximum capacity — the recording remains accessible through the official AWS Summit New York City page.
The three products at the center of today’s session — Kiro, Amazon Bedrock AgentCore, and Amazon Quick — are best understood as a vertical stack rather than isolated announcements. Kiro handles how agents are built and specified. AgentCore handles how those agents run, remember, authenticate, and integrate once deployed. Amazon Quick handles how those capabilities reach knowledge workers and business users. The largest structural implication of this stack: AWS has shipped a specification-to-production pipeline for agentic AI whose foundational design borrows its requirements methodology from aerospace safety engineering.
Read more: AWS Summit New York 2026 Opens Tomorrow With Kiro, AgentCore, and Amazon Quick Reveals
The term “vibe coding” — coined by AI researcher Andrej Karpathy in February 2025 to describe the practice of accepting AI-generated code without reviewing or formally documenting it — became Collins Dictionary’s word of the year for 2025, a distinction that captures both how widespread the habit became and how uncomfortable the engineering community grew with its consequences. By February 2026, Karpathy himself proposed a successor term, “agentic engineering,” describing a more disciplined model in which multiple AI agents handle implementation tasks while human engineers focus on architecture, quality assurance, and strategic decisions. Kiro, AWS’s new agentic IDE launched internationally on May 7, 2026, as the replacement for Amazon Q Developer, is AWS’s production-grade implementation of that shift.
Built on Code OSS — the MIT-licensed open-source base from which Visual Studio Code is derived — Kiro’s defining architectural departure from Q Developer and most other AI coding assistants is that it structurally prevents code generation from beginning until a formal specification exists. When a developer starts a new feature in Kiro, the IDE does not produce code. Instead, it runs through a three-phase workflow that generates three structured documents: a requirements.md file, a design.md file, and a tasks.md file.
The requirements phase is where Kiro’s most distinctive technical decision becomes visible. The requirements.md file is written using EARS (Easy Approach to Requirements Syntax), a formal requirements notation developed by Alistair Mavin and colleagues at Rolls-Royce in 2009, originally to extract airworthiness requirements for a jet engine control system. EARS constrains natural-language requirements into five structural templates using a consistent keyword vocabulary, eliminating ambiguity, vagueness, and incompleteness by imposing clause order on every requirement. Organizations including Airbus, NASA, Intel, and Siemens have adopted it for safety-critical systems engineering. Kiro is the first commercial software IDE to adopt EARS as its default requirements format, importing a specification discipline built for airworthiness certification into everyday software development.
Each requirement in requirements.md follows the pattern: WHEN [condition/event] THE SYSTEM SHALL [expected behavior]. Once requirements are approved, Kiro produces design.md — documenting system architecture, component interactions, and sequence diagrams — and then tasks.md, a dependency-mapped implementation plan. The implementation phase does not proceed linearly. Kiro builds a dependency graph from the tasks in tasks.md and groups independent tasks into concurrent execution waves: Wave 1 contains all tasks with no dependencies and runs them simultaneously, Wave 2 runs tasks whose dependencies were satisfied by Wave 1, and so on until all tasks are complete. This wave-based parallel execution is what gives Kiro’s agent mode its structural efficiency advantage over iterative prompt-and-accept workflows.
Persistent project context is maintained through steering files — markdown documents in a .kiro/steering/ directory that carry the team’s tech stack, file structure, and coding conventions across sessions. Agent hooks allow developers to define event-driven automations that trigger AI agents in the background: a hook can, for example, update test files every time a component is saved, or scan for secrets before a commit.
AWS also confirmed the availability of Kiro Pro Max, a tier designed for professional developers who need sustained, high-volume use across coding, specification generation, and agent-driven tasks, with higher usage limits and access to the latest frontier models. Kiro is currently powered by Claude models accessed through Amazon Bedrock.
One limitation developers have identified in early adoption: because requirements.md and design.md are generated upfront, they do not automatically update when implementation reveals new constraints. Teams have reported specification drift as code evolves beyond what the original spec captured — a “waterfall” dynamic that can slow momentum if not actively managed. AWS has acknowledged the feedback; steering files can partially address it by encoding evolving design constraints that the agent references throughout the session.
Security note: AWS’s security bulletin (CVE-2026-4295) disclosed that versions of Kiro IDE prior to 0.8.0 contained a vulnerability allowing arbitrary code execution when a user opened a maliciously crafted project directory. The issue has been addressed in Kiro IDE version 0.8.0. Developers running earlier versions should upgrade before opening any untrusted project directories.
Amazon Bedrock AgentCore, AWS’s modular enterprise runtime platform for deploying AI agents in production, reached general availability in late 2025 and received expanded capabilities at today’s Summit.
AgentCore’s most important infrastructure characteristic is its session isolation model. Each agent session runs in a dedicated Firecracker microVM — the open-source virtual machine monitor that AWS originally built to power AWS Lambda and AWS Fargate. Firecracker microVMs boot in under 125 milliseconds and impose CPU, memory, and filesystem isolation at the hardware-virtualization level. When an AgentCore session ends, the entire microVM is terminated and its memory is sanitized — no state carries between sessions, and no cross-session data contamination is possible. This is materially different from containerized execution, where namespace isolation can be circumvented by kernel exploits; Firecracker’s hardware-virtualization boundary provides a stronger security guarantee by design.
Sessions can run for up to eight hours, accommodating the long-running, multi-step reasoning processes that complex agentic workflows require. Since June 5, 2026, AgentCore Runtime has supported interactive shell access: a new command opens a PTY-backed terminal directly into the running microVM, with tab completion, color output, Ctrl+C handling, and reconnection on network drop. Each sub-agent in a multi-agent workflow gets its own microVM — isolated filesystems, separate build caches, separate dependencies — with session affinity maintained through a session ID that routes subsequent requests to the same microVM instance.
AgentCore’s component stack is deliberately composable:
For regulated industries, AgentCore inherits Amazon Bedrock’s compliance certifications: HIPAA eligibility, GDPR compliance, SOC 2, ISO 27001, and FedRAMP authorization in GovCloud.
Capabilities expanded at and immediately before today’s Summit include optimization tools — recommendations, batch evaluations, and A/B tests — that close the observe-evaluate-improve loop for agents running in production. Managed payment capabilities, built in partnership with Coinbase and Stripe, are now in preview, enabling agents to autonomously pay for APIs, MCP servers, and web content.
👉 Read more: 318481 GitHub AI agent crisis forces Microsoft to tap AWS
Amazon Quick, the successor to Amazon Q Business, was formally introduced at AWS’s “What’s Next with AWS” event on April 28, 2026. Today’s Summit brought it in front of a broader enterprise audience with expanded capability announcements.
Quick is positioned as an agentic AI assistant for knowledge workers — one that connects to an organization’s data across AWS services, third-party platforms, and on-premise systems, and takes action on that data rather than simply retrieving and summarizing it. The April 28 launch introduced a desktop application for macOS and Windows (in preview), the ability to generate documents, presentations, infographics, and images directly from the chat interface, and native integrations expanding to include Google Workspace, Zoom, Airtable, Dropbox, and Microsoft Teams. Quick now supports more than 100 integrations in total, including Microsoft 365, Salesforce, and Jira. Existing Q Business customers can continue using their current service or migrate their existing Q index to Quick to access the new agentic capabilities without re-ingesting their data.
A notable element of today’s keynote was a case study from Southwest Airlines, represented by Lauren Woods, Executive Vice President and Chief Information Officer. The presentation offered a look at how the carrier has moved agentic AI from proof-of-concept into live operational deployment. Production case studies at hyperscaler conferences tend to be carefully managed; Southwest’s appearance at the Javits Center served as AWS’s most direct evidence that its agentic stack is running in enterprise production today, not only in controlled pilots.
For any team currently using Amazon Q Developer, the transition carries near-term obligations. New Q Developer signups have been blocked since May 15, 2026. Claude Opus 4.6 and all subsequent frontier models are now exclusive to Kiro — teams remaining on Q Developer are locked to an older model stack. Full end-of-support for Q Developer IDE plugins hits April 30, 2027. AWS has committed to critical bug fixes during the transition window, and Kiro’s CLI ships with backward compatibility for Q Developer CLI commands so that existing automation and scripts continue to work during migration.
For enterprise teams on Visual Studio or Eclipse, the migration path to Kiro is more complex than for VS Code users, since Kiro’s current release is built on Code OSS. AWS and practitioners recommend beginning evaluation and parallel running no later than the third quarter of 2026 to leave sufficient runway before the April 30, 2027 deadline.
What is spec-driven development, and why is AWS requiring it in Kiro?
Spec-driven development means generating a formal specification — defining what a system must do and how it must be architected — before any code is written. Kiro requires this because AI coding tools that generate code from natural-language prompts without a specification phase produce code that works in isolation but drifts from architectural intent, misses edge cases, and breaks under production load. By requiring a requirements.md, design.md, and tasks.md before any code generation begins, Kiro makes it structurally difficult to skip the planning phase, regardless of how experienced or time-pressured the developer is.
What is Amazon Bedrock AgentCore’s Firecracker microVM, and why does it matter for security?
Firecracker is an open-source virtual machine monitor originally built by AWS to power AWS Lambda and AWS Fargate. It provisions lightweight virtual machines that boot in under 125 milliseconds and impose hardware-level CPU, memory, and filesystem isolation between sessions. AgentCore uses Firecracker to give each AI agent session its own isolated microVM — so one agent’s code execution cannot access another session’s data, and when a session ends, the entire virtual machine is terminated and its memory is wiped. This is a materially stronger security boundary than containerization alone, which is why AWS has adopted it for enterprise workloads in regulated industries including healthcare, financial services, and government.
When does Amazon Q Developer reach end of support, and what should teams do now?
Amazon Q Developer’s IDE plugins and paid subscriptions reach end of support on April 30, 2027. New signups are already blocked as of May 15, 2026, and the latest Claude models are now available exclusively on Kiro. Teams should begin evaluating Kiro and running it in parallel with existing Q Developer workflows no later than the third quarter of 2026, particularly if they rely on non-VS Code editors such as Visual Studio or Eclipse, where the migration path is more involved.
What is Amazon Quick and how is it different from Amazon Q Business?
Amazon Quick, launched April 28, 2026, is the successor to Amazon Q Business. Where Q Business was primarily a reactive enterprise search and question-answering tool, Quick is designed as an agentic assistant that takes action on your data — generating documents, executing workflows, and integrating with more than 100 business applications including Microsoft 365, Google Workspace, Salesforce, and Slack. Existing Q Business customers can migrate their existing data index to Quick to access the new agentic capabilities without re-ingesting their data.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
Leave a Reply