Source: hubspot.engineering/blog/aviator-hubspots-agent-harness
HubSpot started looking for a shared agent framework in June 2025 because agent implementations across teams varied in features, support and reliability. Most of its backend is Java, and existing tooling was mostly in TypeScript and Python, so the company built Aviator to let teams write agents inside the services they already own. The post says an agent can be written in under 20 lines of Java, and that existing RPCs can be exposed as tools through an annotation-based approach.
Aviator uses a provider-agnostic transcript format, and model clients translate it to each vendor's API. HubSpot says conversations can switch models mid-stream, for example to fall back during an outage. The serialized transcript also works as durable state, so an agent can pause for human approval and resume later on another machine.
To handle large tool counts, Aviator supports deferred tools that stay out of the initial context and are found through tool search, using hosted search where Anthropic and OpenAI provide it and an internal index for self-hosted models. Skills follow the SKILL.md format with extra load-tools and suggested-tools metadata. Code runs in an embedded Python environment and a sandboxed Bash environment, both running in WASM, with microVMs available for some agents.
Why it matters if you build agents
Developers building agents in Java or on similar service stacks can look at Aviator's patterns for transcripts, deferred tool search and skill metadata as a reference design. The post describes these as internal HubSpot systems, so they are not something external developers can install from it.
Teams running multi-model agents can take note of the approach of rewriting tool-call IDs and dropping incompatible reasoning blocks when switching providers. Anyone storing transcripts should plan for the fact that HubSpot treats the transcript as the full conversation state.
Get stories like this every weekday
One short email with the day's agent tooling news.