HubSpot built Aviator as a shared Java framework for creating and operating AI agents inside the company’s existing services. Before it, teams made inconsistent choices about models, tool calling, and context management, often repeating mistakes that made agents slower, costlier, and less reliable. Existing Python frameworks were difficult to integrate with HubSpot’s Java backend, so the team set a goal of making an agent possible in fewer than 20 lines of Java. Aviator lets developers reuse existing CHIRP RPCs as agent tools; an infrastructure team used its RPCs to build a debugging agent in about 15 minutes.
Aviator is designed to keep agents adaptable and manageable at scale. It stores conversations in a provider-neutral transcript format and translates them for different model APIs, allowing a conversation to switch models, resume later, or pass its context to another agent. The serialized transcript also supports durable checkpoints and debugging traces. To handle agents with thousands of available tools, Aviator can defer tool definitions and expose them through search. Its SKILL.md support packages task instructions with tools, references, and scripts, loading details only when needed. Together, these features aim to make improvements to reliability, model choice, and tool use available across teams rather than leaving each agent builder to solve the same problems independently.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.