HubSpot uses AI agents to test large-scale code rewrites

Dharmesh Shah describes how agentic coding allows established software companies to undertake large-scale architectural experiments that were previously cost-prohibitive. He explains that HubSpot can use AI models to rewrite significant portions of its codebase to test different architectural choices, such as changing data structures or languages, effectively exploring 'alternative universes' of technical decisions. This process is validated using existing unit tests, now referred to as evals, to ensure the rewritten code maintains functionality.

Shah notes that this capability allows companies to revisit and optimize primitives built years ago, leveraging the speed of AI to iterate on complex engineering problems. He also highlights the importance of maintaining a 'system of record' for structured data alongside a 'system of context' for unstructured data, arguing that deterministic queries should be used for known facts rather than inference. Shah argues that AI coding agents enable incumbents to perform deep, large-scale architectural refactoring and testing, turning historical technical debt into an asset for innovation.