AWS re:Invent 2025 - Extend Kiro with MCP support for richer context (DVT213)

Extending Kiro with MCP: Transforming Agent-Driven Development

Context is Everything for Effective AI Agents

• The quality of an AI agent's outputs is directly tied to the quality of the context it is provided
• Context engineering is the critical step that separates prototype AI systems from production-ready solutions
• Context includes:
  • System context (role definition, behavioral instructions)
  • Retrieved context (knowledge from indexed sources)
  • Conversational context (previous messages and actions)
  • Tool context (dynamic information from integrated tools and APIs)

Limitations of the Retrieval-Augmentation-Generation (RAG) Approach

• RAG automates the process of retrieving relevant information from knowledge bases to augment the user's prompt
• While effective for some use cases, RAG has key limitations:
  • Passive system - cannot request additional information or clarification
  • Semantic gap - retrieved information may not match the agent's actual needs
  • Context growth - message history can quickly overwhelm the context window

Introducing Model Context Protocol (MCP)

• MCP is an open, client-server protocol for connecting AI agents to data and tools
• Key MCP primitives:
  • Resources: Expose data sources and APIs
  • Prompts: Request information or actions
  • Tools: Execute external functionality
• MCP enables a closed feedback loop where agents can:
  • Propose solutions
  • Take actions (edit code, launch servers, etc.)
  • Observe outcomes
  • Iterate on solutions

Leveraging MCP in Kiro's Development Lifecycle

• Kiro team used MCP across the entire software development lifecycle:
  • Ideation: Retrieve backlog, designs, and requirements
  • Implementation: Access reference docs, control dev servers, test changes
  • Testing: Benchmark agent performance, conduct regression analysis
  • Deployment: Diagnose issues by retrieving logs and telemetry
• MCP enables agents to be more autonomous and self-sufficient

Addressing MCP's Context Management Challenges

• MCP's rich integration capabilities can quickly consume an agent's available context
• Kiro introduces "Powers" - pre-configured MCP integrations that provide just-in-time context
• Powers allow agents to dynamically load only the necessary context, improving efficiency

Key Takeaways

• Context engineering is critical for production-ready AI agents
• MCP enables a closed feedback loop where agents can propose, act, observe, and iterate
• Kiro leveraged MCP across the entire software development lifecycle
• Kiro's "Powers" feature addresses MCP's context management challenges

Technical Details

• MCP supports two transport mechanisms: stdio (local) and streamable HTTP (production)
• Kiro's MCP integrations include AWS Docs, Chrome, GitHub, and Figma
• Kiro uses MCP to automate tasks like launching dev servers, running benchmarks, and diagnosing production issues

Business Impact

• Kiro's use of MCP and agent-driven development accelerates software delivery by automating repetitive tasks
• Agents can quickly onboard, diagnose issues, and iterate on solutions, improving overall productivity
• The "Powers" feature ensures efficient context management, enabling agents to focus on high-value work

Example: Resolving a 2048 Game Bug

• Kiro agent used MCP to:
  • Verify the reported bug by interacting with the local 2048 game
  • Analyze the source code to propose a fix
  • Validate the fix by reloading the game and confirming the behavior
• The agent then updated the GitHub issue with its findings and proposed solution