Amazon Web Services Inc. today introduced a new feature for Kiro, the company’s artificial intelligence-powered agentic coding environment for software developers, adding new “powers” that bring specialized expertise to the tool.

These powers, unveiled at the company’s re:Invent conference in Las Vegas, load only when needed, providing on-demand augmentation without occupying the model’s context window unnecessarily.

Kiro is Amazon’s AI-powered agentic coding environment designed to help developers write, analyze and build software using natural language prompts. The company launched Kiro into general availability last month. The software sets itself apart from current “vibe coding” options on the market by first producing a “spec,” or specification, which plans out clear requirements and structured designs before proceeding.

When using an AI agent for coding, developers must balance the number of tools they activate with how much upfront information they feed the system. Too much information and agents consume tokens faster, making runs more expensive. Tool definitions also take up context, or running “memory,” that could otherwise be used to hold code or assist reasoning. Load too many tools at once and everything bogs down.

Kiro powers address this by providing a unified way to dynamically load tools, Model Context Protocol configurations and custom instructions as needed, rather than all at once. The existing ecosystem already allows for dynamic loading — from MCP servers to skills — but configuration has traditionally been fragmented across multiple files and patterns.

Imagine an agentic coding environment that automatically loads skills on demand when a user asks for them. Mention “database” and powers detect an installed Neon configuration, load it and run with it. Install the Stripe power and mention “payment” or “checkout,” and Kiro will activate that capability to generate code using Stripe’s application programming interfaces.

With each task change, the previously loaded power deactivates, freeing up context for the next operation. This keeps the model’s running memory uncluttered while allowing agents to focus on the job at hand.

MCP servers standardize how AI agents access external tools and data. They’re especially powerful for agentic coding because they replace one-off, custom integrations with a universal protocol, letting agents tap into real-world context and interact with third-party systems as needed. In practice, an MCP server can expose capabilities such as running tests, debugging code or generating the specific scaffolding required to connect to an external service.

Traditional MCP servers load all of their tools up front. A Figma server, for example, might expose eight tools requiring around 12,000 tokens. A Postman server might add more than 120 tools. Loading everything all at once can quickly consume a massive portion of context that could otherwise be allocated to reasoning or code.

Developers can browse and install powers directly within Kiro or via kiro.dev. Once installed, powers register themselves and gain their full configuration. Amazon is launching with a broad list of partners, including Figma Inc., Stripe Inc., Neon Inc., Supabase Inc., Netlify Inc. and Datadog Inc. Users can also import powers from GitHub for community-built tooling, or privately develop their own for internal use.

At launch, powers work only inside the Kiro development environment, but Amazon said it intends to make them compatible with any AI development tool. Soon, powers will be available within the Kiro command line, as well as Cline, Cursor, Claude Code and other AI-assisted programming environments.

Image: SiliconANGLE/Microsoft Designer