AWS re:Invent 2025 - Accelerate development with Kiro's new agentic AI capabilities (DVT228)

Source: Dev.to

Overview

AWS re:Invent 2025 – Accelerate development with Kiro’s new agentic AI capabilities (DVT228)

Kiro Powers is a new feature that bundles MCP (Model Context Protocol) and steering instructions to give AI agents specialized capabilities without overwhelming them with context. The presentation demonstrates Powers through live demos with Figma, Postman, and Supabase integrations.

• Harry Mauer (CTO, Postman) shows how a Power can automatically create API collections and mock servers directly in the IDE.
• Tyler Schuckert (DevRel Engineer, Supabase) builds a complete personal‑banking app, covering database schema, security checks, TypeScript type generation, and authentication.

Powers can be installed with a single click from a curated marketplace or custom‑built using a power.md file, keeping agents extensible while preserving performance through progressive disclosure of context.

This article is auto‑generated from the original presentation; minor typos may remain.

---

Introduction to Kiro Powers: Addressing the Generalist Problem in AI Agents

Kiro Powers tackles the “generalist” limitation of AI agents. Instead of feeding an agent massive amounts of context for every task, Powers bundle the necessary MCP definitions and steering instructions, enabling agents to act as specialists on demand.

Agenda (as presented)

1. The general problem we aimed to solve with Powers.
2. How Kiro Powers address that problem.
3. Live demos with our partners:
   • Harry Mauer – Postman (automatic API collection creation & mock server setup).
   • Tyler Schuckert – Supabase (end‑to‑end personal‑banking app build).

The Challenges of Specialized Tasks: From MCP to Context Rot

Generalist AI agents

AI has moved from inline code suggestions to full‑blown agents capable of building large features. However, agents remain generalists: they require extensive context to perform specific tasks (e.g., front‑end work, API design), leading to cumbersome prompting and inconsistent results.

Rapidly evolving tooling

Software ecosystems evolve daily—new SDKs, frameworks, and libraries appear constantly. Models powering agents are often trained on outdated data, so they may lack knowledge of the latest tools (e.g., a newly released TypeScript SDK). Simply linking documentation is insufficient; agents may still make incorrect assumptions.

Model Context Protocol (MCP)

MCP is a convention that lets agents interact with external services via a structured API, enabling actions such as listing projects, updating tables, or creating collections. Both Postman and Supabase expose MCP servers that the agents can call.

Context rot

Adding many MCPs or tools to an agent can overwhelm it, causing context rot—performance degradation and lower‑quality output as the agent’s context window fills with irrelevant data. Moreover, agents may miss onboarding steps or guidance that are not encoded in the MCP alone.

Existing alternatives

• Claude Skills (Anthropic) – dynamic tool discovery.
• Cursor rules, agents.md, sub‑agents, etc.

These solutions aim to solve the same problem—providing agents with focused, up‑to‑date capabilities—yet the ecosystem is fragmented, leaving developers confused about which approach to adopt.

Goal

Provide a unified mechanism (Kiro Powers) that combines MCP‑based tool access with concise steering instructions, delivering specialist behavior without incurring context rot.

Demo Highlights

Postman Power (Harry Mauer)

• One‑click creation of API collections.
• Automatic mock server provisioning directly from the IDE.

Supabase Power (Tyler Schuckert)

• Builds a personal‑banking app end‑to‑end.
• Generates database schema, enforces security checks, produces TypeScript types, and configures authentication—all via a single Power.

Getting Started with Kiro Powers

• Marketplace install – Browse the curated Power marketplace and install a Power with a single click.
• Custom Power – Create a power.md file defining MCP endpoints and steering instructions, then load it into your agent.

Both methods maintain performance by exposing only the necessary context when the Power is invoked.

For the full video, visit the YouTube playlist linked in the thumbnails above.