Block's Buzz puts staff and AI agents in one workspace

Block's move reflects a practical problem that arises once AI agents become common inside a company: employees end up juggling separate tools and models for coding, analytics, and everyday requests. Buzz addresses that by making agents participants in the same conversations where colleagues already work, so an agent's output can be reviewed and extended by other people in place. The choice to make Buzz open source and model-agnostic lowers the barrier for other organizations to inspect or adapt it, and self-hosting gives a path for running it privately.

The Databricks Unity Gateway integration is the piece that makes model choice practical at scale. Rather than wiring each agent to a separate provider, employees select Databricks as their inference provider and pick from models available through the gateway, including frontier models from OpenAI, Anthropic, and Google and select open-weight models. That setup appears designed to let teams build a collection of specialized agents rather than standardizing on a single model.

Block's internal experiments point to where this is heading. Engineers have tried having a more capable model coordinate a team of agents running on smaller, more efficient models, on the reasoning that not every step of a complex task needs the most expensive model. Block also built an internal benchmarking approach for these agent teams. The company frames the question as shifting from which single model performs best to how models, harnesses, and agents combine to complete work.