Al Harris is a principal engineer at Amazon Web Services and an early technical leader behind Kiro, the AI development platform built around spec-driven development. He builds tools that translate a developer’s intentions into requirements, technical designs, implementation tasks, and tests, making AI-generated software easier to inspect and verify.
Harris helped develop Kiro from an experimental Visual Studio Code extension into an agentic development environment. Initial experiments emphasized test-driven workflows; internal feedback expanded the product’s scope to encompass requirements, design, implementation, and verification. He has since contributed to Kiro’s IDE diagnostics, informed its approach to safer semantic refactoring, and helped design the shared infrastructure connecting its desktop, command-line, web, and mobile experiences.
- Specifications as living documentation. Harris treats specifications as evolving records of software behavior and architectural decisions, not disposable execution plans. Teams can review requirements collaboratively, revise acceptance criteria as systems change, and trace those changes into implementation.
- Requirements connected to verification. Structured acceptance criteria can become explicit correctness properties, allowing property-based testing to search for counterexamples. Harris also favors automated reasoning to identify ambiguous or contradictory requirements before code is written.
- Agents grounded in actual engineering context. Through the Model Context Protocol, agents can incorporate project-management tasks, technical documentation, and other external information throughout development. His Lofty Views sample application demonstrates that workflow. Persistent steering documents capture practical conventions such as deployment procedures, test expectations, and attribution for generated commits.
- Selective context and dependable tools. Harris favors incremental context discovery over overwhelming agents with entire codebases, and established language-server diagnostics and semantic refactoring over probabilistic guesses. His AI Engineer Code demonstration also illustrates how prematurely specifying Amazon S3 can narrow an agent’s architectural reasoning before better persistence options are considered.
In 2026, Harris coauthored the architecture behind Kiro’s unified agent harness, replacing separate TypeScript, Rust, and Python agent implementations with a standalone server communicating through the Agent Client Protocol. The design shares orchestration, permissions, tools, and session behavior across interfaces while preserving each client’s native experience.
