The closing session of Agent Sessions Day featured James Montemagno interviewing Kai Maetzel about the evolution and future of AI agents in Visual Studio Code (VS Code). They discussed how the rapid advancements in AI models have significantly shifted the way developers interact with code, moving from traditional manual coding to more agent-driven workflows. Kai emphasized that this trend is only going to accelerate, with agentic coding becoming a primary activity for more developers, not just a select few. As agents become more capable, the focus is shifting from just code editing to managing agent sessions and workflows within VS Code.
Kai outlined a vision where agents are integrated throughout the entire software development lifecycle, not just in planning or coding, but also in reviewing, testing, and validation. He highlighted the importance of customizing workflows to fit individual or team preferences, such as determining when code is ready to be auto-committed or what level of testing is required. The infrastructure for building custom agents is already available, and Kai expects the community to share best practices and combine different agents for optimal results, much like assembling a team of specialists.
A key point discussed was the orchestration of multiple agents and models, each with unique strengths. Kai and James noted that using different models (like Gemini, Codex, or Claude) for code review can yield diverse perspectives, similar to having multiple human reviewers. This approach leverages the unique qualities of each model, such as Codex’s strength in performance analysis, to improve code quality. The flexibility to use the best model for each task, supported by GitHub’s multi-model capabilities, is seen as a major advantage for developers.
The conversation also touched on VS Code’s commitment to openness and extensibility in the AI-first era. Kai reaffirmed that open source principles and community contributions remain central, with features like open-sourced extensions, bring-your-own-key providers, and customizable agent loops. The extensibility model has evolved beyond traditional extensions to include custom agents, hooks, and MCP servers, allowing developers to tailor their workflows and even create interactive UI elements within chat experiences. This shift enables more granular and workflow-specific customization without always needing to build full extensions.
Finally, Kai shared insights into how the VS Code team itself has adapted to these changes, using agentic flows internally to speed up development and experimentation. The team now spends less time on implementation and more on ideation and user feedback, as new models and tools make execution faster. Kai’s pro tip for users is to embrace customization and incremental development, using agents to prototype, generate specs, and refine workflows. He encourages developers to focus on what brings them satisfaction and leverage agentic tools to handle the rest, reflecting the broader shift towards more human-centric, flexible, and efficient software development.
