AWS re:Invent 2025 - Teaching your AI tools custom languages and libraries (DVT341)

Summary of AWS re:Invent 2025 Presentation: "Teaching your AI tools custom languages and libraries (DVT341)"

Introduction to Kira - AWS's Agentic IDE

• Kira is an "agentic IDE" launched by AWS in 2025, building on previous AI-powered developer tools like Code Whisperer and Amazon Q.
• Kira takes a goal-oriented, reasoning-based approach to assist developers throughout the entire software development lifecycle.
• Key capabilities of Kira include:
  • Prototype to production support
  • Integrating with external systems via the Model Context Protocol (MCP)
  • Customizing Kira's behavior through "steering" - rules and preferences defined by the developer

Challenges of Custom Languages and Libraries

• Many organizations have custom tools, languages, and libraries that are not known to large language models.
• Examples include:
  • Simple custom logging libraries
  • Entire custom technology stacks
  • Domain-specific languages (DSLs) for business logic and workflows
• These custom elements pose a challenge for AI assistants like Kira, which need to be taught how to understand and work with them.

Introducing "Math JSON" - A Fictional DSL

• For the purposes of the demo, the presenter introduces a fictional DSL called "Math JSON" - a JSON-based syntax for representing mathematical equations.
• Math JSON is completely made up, but serves as a proxy for the types of custom languages and formats that real-world organizations often use.
• The presentation walks through a series of demos showing how to teach Kira to work with Math JSON, starting from a baseline of Kira having no knowledge of the language.

Iterative Approach to Teaching Kira Custom Languages

1. Baseline: Kira initially fails to understand Math JSON, defaulting to generating Python code instead.

2. Steering: The presenter adds a "steering" file to Kira's context, providing guidance on how to work with Math JSON, including file extensions, linting, and execution.

3. Refining Guidance: Kira is asked to refine the steering documentation, condensing it and removing redundant information.

4. Integrating with MCP: Instead of maintaining a local copy of the Math JSON documentation, Kira is configured to fetch the documentation from a remote source (a GitHub repository) using the Model Context Protocol (MCP).

Key Takeaways

• AI assistants like Kira need to be explicitly taught about custom languages, libraries, and tools used within an organization.
• This can be achieved through a combination of "steering" (providing guidance on preferred practices) and integrating with external data sources via MCP.
• Kira can be empowered to take an active role in refining and improving the guidance provided to it, optimizing the information it needs to work effectively with custom elements.
• Integrating with remote documentation sources, rather than maintaining local copies, helps ensure the information Kira uses stays up-to-date.

Business Impact and Real-World Applications

• The ability to teach AI assistants like Kira about custom languages and libraries enables organizations to leverage the full power of these tools across their entire codebase and technology stack.
• This can lead to increased developer productivity, reduced maintenance overhead, and better alignment between AI-powered tools and an organization's unique technical environment.
• The techniques demonstrated can be applied to a wide range of custom elements, from simple libraries to complex domain-specific languages used for business workflows and logic.

Example Use Cases

• Onboarding new developers by providing Kira with the necessary information to understand an organization's custom tools and languages.
• Automating the maintenance of custom libraries and DSLs by having Kira proactively update its own guidance based on changes to the source documentation.
• Leveraging Kira's capabilities to generate code, perform analysis, and execute tests against custom languages and libraries, ensuring they continue to function as expected.