Draft:Kiro

Motivation

Kiro was introduced as an AI-assisted development environment focused on “spec-driven development,” which its creators presented as a response to limitations they associated with “vibe coding.” In launch materials and early coverage, Kiro was presented as addressing cases where prompt-driven coding can leave software harder to document, review, and maintain in larger or production-oriented projects. In the launch announcement, Kiro’s creators argued that AI-assisted coding workflows can leave requirements unclear, decisions undocumented, and system design harder to understand, while later maintenance suffers when knowledge is not captured in shared artifacts.

In the launch post, Nikhil Swaminathan and Deepak Singh said Kiro was intended to improve design alignment, requirements clarity, review rigor, and long-term maintainability in AI-assisted software development.^[1]

Early coverage described Kiro as aiming to bridge rapid AI prototyping and production-ready software; GeekWire highlighted the company’s tagline, “from vibe coding to viable code.” ^[2]

Announcement and launch

Kiro entered public preview on July 14, 2025. AWS Summit New York City took place on July 16, 2025. In a post on Twitter / X, Amazon CEO Andy Jassy said Kiro “has a chance to transform how developers build software,” and pointed to spec-driven development, intelligent agent hooks, and a purpose-built interface for developer workflows..^[3]

Early coverage contrasted Kiro with Amazon Q Developer, describing Q Developer as focused on code completion and chat-based assistance, while Kiro was presented as a standalone IDE built around more autonomous agent workflows. Unlike Q Developer, Kiro was designed as a full standalone IDE with a separate brand identity. DevClass reported that AWS developer advocate Nathan Peck said Kiro was intended to have a “unique identity outside AWS” in order to appeal to developers using other platforms. Kiro does not require an AWS account; users can sign in with Google, GitHub, AWS Builder ID, or AWS IAM Identity Center.^[4]

Kiro reached general availability on November 17, 2025. The general-availability release added property-based testing, checkpointing, multi-root workspace support, the Kiro CLI, and enterprise billing.^[5]

Code OSS base

Kiro is built on Code OSS, the MIT-licensed open-source core of Visual Studio Code, and is compatible with Open VSX registry plugins. This makes it one of several AI-focused IDEs in a growing market of VS Code forks, alongside Cursor, Windsurf, Trae, PearAI, and Void. Because it uses Open VSX rather than Microsoft's proprietary extension marketplace, some Microsoft-exclusive extensions — such as the official C# extension — are not available in Kiro, which has been noted as a limitation for .NET developers.^[6]

AI models

Kiro uses AI models developed by Anthropic, primarily members of the Claude family. As of April 2026, supported models include Claude Sonnet 4.6 and Claude Opus 4.6 — both of which support a 1 million token context window and are generally available on Pro, Pro+, and Power tiers — as well as older Claude Sonnet and Opus versions. Kiro has also added experimental support for third-party open-weight models, including MiniMax M2.5 (with a 200K context window) and GLM-5, a model suited for large-codebase migrations and cross-file reasoning. All models can be selected from the Kiro model selector inside both the IDE and CLI.^[7]

Enterprise administrators can restrict which models are available to their organization, which is particularly relevant for data residency requirements where models using global cross-region inference may need to be excluded until regional inference becomes generally available.

Language support

Kiro supports a broad range of programming languages, including Python, Java, JavaScript, TypeScript, C#, Go, Rust, PHP, Ruby, Kotlin, C, C++, shell scripting, SQL, Scala, JSON, YAML, and HCL, among others.^[8]

Spec-driven development

Spec-driven development is Kiro's primary differentiating feature. Rather than generating code immediately in response to a natural language prompt, Kiro first produces a set of structured planning documents — collectively called a spec — before writing any code. A spec consists of three markdown files:

• requirements.md — A requirements document written using EARS notation (Easy Approach to Requirements Syntax), a method for writing unambiguous textual requirements originally developed at Rolls-Royce for safety-critical aerospace systems. Requirements are expressed as user stories with detailed acceptance criteria covering edge cases.^[9]
• design.md — A system design and architecture document that records technology stack selections, component interactions, data flows, and architectural decisions made to fulfill the requirements.
• tasks.md — A dependency-ordered implementation checklist with concrete tasks that AI agents execute sequentially, including optional comprehensive tests.

These documents serve as a durable source of truth for both human developers and AI agents. Analysts at Medium and Builder.io have identified Kiro as the only major AI coding tool as of 2026 that produces "an auditable trail from requirement to implementation," making it particularly valuable for teams in regulated industries that require traceability.^[10][11]

Agent modes

Kiro offers two modes for controlling how autonomously the AI agent operates:

• Autopilot mode — The agent operates autonomously, making multiple changes across the codebase without requiring approval for each action. Developers can review a comprehensive diff of all changes through a "View all changes" option in the chat interface. Autopilot mode is the default setting and is recommended for well-understood tasks such as adding tests, aligning documentation, or generating boilerplate code.
• Supervised mode — The agent pauses after each turn that involves file edits and presents changes as individual hunks (as of version 0.10), allowing developers to review and approve each change before it is applied. Supervised mode is recommended for critical components such as authentication systems, infrastructure code, or anything that touches external contracts.

Both modes work in both vibe chat and spec chat sessions. Developers can toggle between modes at any time from the chat interface.^[13]

Agentic chat (Vibe mode)

In addition to its structured spec-driven workflow, Kiro includes a free-form conversational coding mode — colloquially referred to as vibe mode or agentic chat — for quick, ad-hoc coding tasks where the full spec workflow would be disproportionate. This mode supports multimodal inputs, including code files, images, URLs, terminal output, and documentation. As of early 2026, users can also attach documents directly to chat messages in formats including PDF, CSV, DOCX, XLSX, HTML, TXT, and Markdown, with up to five documents per message.^[14]

Hooks

Kiro Hooks are event-driven automations that trigger background agent tasks in response to developer actions. Hooks can be configured to fire on events such as saving a file, creating a new file, modifying an API endpoint, or making a commit to a repository. As of version 0.10, hooks can also trigger on spec task events. Common hook use cases include:

• Automatically generating unit tests when source files are modified
• Refreshing documentation when API routes change
• Enforcing naming conventions on new file creation
• Running lint or security scans on specific file paths
• Blocking modifications to protected files

Kiro also supports manually-triggered hooks using the userTriggered hook type, which can be fired on demand for specific workflows. Amazon CEO Andy Jassy described hooks as "like having an experienced developer constantly reviewing your work and handling the maintenance tasks that often get delayed."^[15]

Steering files

Developers can define persistent steering files — stored as markdown files in a .kiro/steering/ directory — that encode project-wide behavioral guidelines for the AI agent. Steering files persist across sessions and can capture information such as preferred tech stack, code style conventions, architectural patterns, security requirements, or framework-specific rules. The agent applies these rules consistently without the developer needing to re-specify them in each conversation.

Checkpointing

Kiro supports checkpointing, which allows developers to save the state of a conversation and revert the workspace to a previous checkpoint at any time. This enables developers to undo AI agent changes at the conversation level — rather than line by line — and return the filesystem to a prior state, with guidance on next steps provided in the interface.^[16]

Property-based testing

Kiro includes property-based testing (PBT) support, introduced at general availability. Unlike standard unit tests, property-based tests run hundreds of times with randomly generated inputs to detect edge cases that specific, manually chosen test values would not reveal. Kiro generates property-based tests from the acceptance criteria defined in specification documents, using them as verification evidence for spec requirements.^[17]

Model Context Protocol (MCP) support

Kiro includes native support for the Model Context Protocol (MCP), allowing the agent to connect to external tools and data sources such as documentation servers, API clients, database interfaces, and cloud platform integrations. As of version 0.10, MCP servers can also surface prompts, templates, and elicitation support directly in the Kiro chat interface.

Kiro Powers

Kiro Powers are first-party and third-party extensions that augment the core Kiro agent with purpose-built context and tooling for specific workflows. Powers provide agents with only the tools and context they need for a given task, reducing noise and improving output quality. Powers can be installed directly from within the Kiro IDE or web interface without requiring manual MCP server configuration. Powers launched at AWS re:Invent 2025 with integrations for Figma (design integration) and Netlify (hosting and deployment). A subsequent Power released in February 2026 added AWS IAM Policy Autopilot, enabling developers to generate baseline IAM policies from within the IDE without leaving their coding workflow.^[18]

Kiro CLI

Launched at general availability in November 2025, the Kiro CLI brings Kiro's agentic capabilities to the terminal. It supports agent modes, MCP servers, steering files, and custom agents. As of CLI version 2.0 (early 2026), the CLI expanded to native Windows 11 support and introduced headless mode, enabling non-interactive use in CI/CD pipelines, automation scripts, and environments without a browser. Authentication in headless mode is handled via a KIRO_API_KEY environment variable.

The CLI also supports the Agent Client Protocol (ACP), an open standard that allows Kiro to function as an agent within ACP-compatible editors including JetBrains IDEs and Zed, extending Kiro's spec-driven capabilities beyond its own IDE interface.^[19]

The CLI includes a built-in Plan Agent, accessible via Shift+Tab or the /plan command, which gathers structured requirements through interactive questions, researches the codebase, and produces a detailed implementation plan before handing off to the execution agent. The Plan Agent operates in read-only mode and cannot modify files during the planning phase.

Autonomous agent (Frontier Agent)

At AWS re:Invent 2025, Amazon announced the Kiro Autonomous Agent as part of a new class of AI systems AWS termed "Frontier Agents" — autonomous agents capable of operating independently for extended periods across multiple tasks and repositories. The Autonomous Agent was made available in preview to Kiro Pro, Pro+, and Power subscribers.

Unlike standard agentic coding sessions that reset when closed, the Kiro Autonomous Agent maintains persistent context across sessions, repositories, and pull requests. It learns from developer code review feedback — for example, if a reviewer comments "always use our standard error handling pattern," the agent internalizes that preference and applies it to future work automatically. Key capabilities include:

• Asynchronous task execution in isolated sandbox environments, allowing work to continue while the developer focuses elsewhere
• Multi-repository coordination, letting a single task trigger coordinated edits and pull requests across multiple codebases
• Persistent cross-session memory of team conventions, codebase structure, and past decisions
• Automatic spawning of specialized sub-agents to handle components of a complex task

AWS CEO Matt Garman announced at re:Invent 2025 that Amazon is standardizing its own internal software development teams on Kiro, describing it as "the agentic development environment for structured AI coding."^[20]

Platforms

Kiro is available as a desktop application for macOS, Windows, and Linux. The Kiro CLI is available on macOS and Linux, with native Windows 11 support added in CLI version 2.0. A GovCloud deployment is also available for US government and regulated-industry customers, with inference processed through Amazon Bedrock in AWS GovCloud (US-West). GovCloud authentication is handled exclusively via AWS IAM Identity Center; individual login methods (Google, GitHub, AWS Builder ID) are not available in GovCloud regions.

Pricing tiers

Kiro was offered free of charge during its public preview period (July–November 2025). Following general availability, pricing is credit-based, with overage charged at $0.04 per credit on paid plans. As of early 2026, the public pricing tiers are:

Credits do not roll over between billing periods. New users receive a 30-day trial bundle with 500 bonus credits. Eligible university students receive one full year of free access with 1,000 credits per month. A startup program is available for companies that have raised up to Series B funding, providing AWS credits toward Kiro Pro+.^[21]

Content from paid-tier users is not used to train AI models. Free-tier users may opt out of model training in settings. Telemetry is disabled by default for enterprise users.

Critical and industry analysis

Upon its launch, Kiro received coverage from technology publications including CNBC, The Register, GeekWire, DevClass, TechTarget, TechRadar, and Visual Studio Magazine. The Register characterized the product's approach as a counterpoint to dominant "vibe coding" tools, noting that its spec-driven methodology "will help to mitigate the sometimes-disorderly outcomes from AI-driven coding."^[22]

TechTarget analyst Jason Andersen of Moor Insights & Strategy highlighted spec-driven development and limited AWS lock-in as Kiro's two most distinctive qualities, noting that it is "the first thing [Kiro] asks you when you start a new project: 'Do you want to start with spec, or do you want to start with prompts?'"^[23]

Constellation Research analyst Holger Mueller commented that the product reflects a fundamental change in the field, observing that "software development and coding are not the same anymore in the era of genAI."^[24]

Developer community analysis published in late 2025 identified Kiro, Cursor, and Windsurf as the three dominant forces in the agentic IDE market, each representing a distinct philosophy: Cursor optimizing for speed and iterative developer flow, Windsurf targeting enterprise monorepos, and Kiro targeting structured, spec-first, auditable development.^[25]

RedMonk analyst Kate Holterhoff identified Kiro as an industry benchmark for transparent usage reporting, noting that unlike some competitors whose credit systems caught users off guard, Kiro gives developers granular data on what each prompt costs before limits are reached.^[26]

A comparative review published in March 2026 covering seven major AI coding tools noted that Kiro is "the only tool that creates an auditable trail from requirement to implementation," making it uniquely suited for regulated industries and teams that require compliance traceability, and recommended it specifically for teams where audit trails and compliance are requirements.^[27]

Criticism and limitations

Reviewers and analysts noted several limitations. DevClass observed that Kiro's default Autopilot mode — in which the agent can make code changes autonomously — may concern developers unfamiliar with its behavior.^[28] Some .NET developers noted that the exclusion of Microsoft's proprietary C# extension — unavailable on the Open VSX registry — was a practical friction point.

Credit consumption rates by operation type were criticized as inconsistently published, making cost forecasting difficult for teams at scale. AWS also acknowledged early bugs that caused unexpected credit drainage. Analysts at Augment Code noted that Kiro lacked a public trust center and limited audit documentation as of early 2026, creating challenges for independent compliance evaluation without direct AWS sales engagement.^[29]

Enterprise adoption

At AWS re:Invent 2025, Delta Air Lines presented a detailed case study of its enterprise deployment of Kiro. The airline reported a 1,948% increase in Amazon Q Developer adoption within six months, driven by a trust-building rollout through internal developer champions. Kiro's spec-driven approach was credited with transforming backlog grooming sessions into efficient structured design sessions that non-technical product owners could participate in. Delta reported a 94% developer satisfaction score from pilot users and successfully delivered its first intelligent developer portal using Kiro, with developers shifting their focus from writing implementation code to reviewing and directing architecture.^[30]