Vaibhav Gupta opens with a deliberately provocative account of a team that does not conduct routine code reviews, lets engineers use different AI tools and works in parallel. Rather than endorsing unchecked generated code, he describes replacing some manual coordination with a short, model-agnostic architecture document, versioned design documents, required human readers and automated checks on dependency boundaries. These are the practices he says his team uses while building a programming language.
Vaibhav Gupta then outlines an agent-driven feedback loop: agents generate programs, inspect tool-use transcripts and surface possible language or developer-experience problems; humans decide which findings are real; agents can propose fixes; and alternatives can be compared by tool calls, errors and outcomes. He argues that execution traces and navigable views of code can help engineers understand a system without reading every line. His assertions about team productivity, stable architecture and tracing costs are first-person claims, not independently verified results.
The second half shifts from process to agent-first language and tooling design. Vaibhav Gupta demonstrates ideas for semantic code navigation, runnable functions, portable command-line tools, inferred error types and calls across language boundaries. He argues that compilers and type systems can enforce useful guarantees for both people and agents while allowing adoption from existing languages rather than requiring a complete rewrite. The talk ends by urging engineering teams to adapt their processes to agent-speed development, while leaving open how broadly the demonstrated approach generalizes.
Watch on YouTube
