Block launches Buzz, an AI collaborative space with Unity Gateway

Block operates Buzz, an open-source space where employees and AI agents collaborate. Connected to Databricks Unity Gateway, the tool provides access, through a single integration, to models from OpenAI, Anthropic, Google, and open weights. The company is testing teams of agents coordinated by a more capable model to balance cost and performance.

Combinations of Agents Tested for Cost and Performance

Block has experimented with workflows involving multiple agents to compare the impact of different model-harness pairs on performance and cost. Among the methods evaluated, engineers tried a scheme where a more capable model coordinates agents based on smaller, more efficient models. According to Block's engineers, it is not necessary to use the most expensive model at every step of a complex task. A coordinating agent can handle the most complex segments, while other agents take care of subtasks that match their own capabilities. Block has also developed an internal method to evaluate teams of agents, aiming to analyze different configurations. This approach marks an evolution, favoring the combination of models, harnesses, and agents to perform tasks efficiently, rather than relying on a single model.

A Shared Space Where Humans and Agents Co-Work

Buzz brings together employees and AI agents in a shared workspace to exchange information, add context, and collaborate on tasks. The tool allows for group discussions, focused conversations, and one-on-one exchanges. An employee can, for example, request an agent for code or analysis, while colleagues review, complement, or ask questions. Interactions between agents can occur in view of the participants in this environment, facilitating human intervention in the workflow. Many Block employees thus integrate agents directly into conversations where collaboration with their colleagues is already underway.

Unified Access to Models via Databricks Unity Gateway

Within Block, Buzz connects to Databricks Unity Gateway to offer, through a single integration, access to a wide range of models. When creating an agent, users have the option to choose Databricks as the inference service and select models from OpenAI, Anthropic, Google, or certain open-weight models. This centralized access interface makes it possible to test and adapt the chosen model according to task needs. This flexibility is deemed important as new models emerge and teams identify the most suitable ones.

Creation of Specialized Agents and Task-Based Selection

Buzz allows users to control the model or harness underlying each agent, depending on the nature of the task. Rather than imposing a single model, the approach enables the composition of a collection of specialized agents, each optimized for a specific use, and to bring them together in the same workflow. Teams do not have to switch between distinct tools to collaborate with multiple agents. At Block, this flexibility is leveraged for coding, analysis, and other daily needs, responding to a demand for collaboration with agents in the same context as regular work.

Open Source, Agnostic, and Self-Hostable

Buzz includes a native agent based on a lightweight version of Goose and supports other coding harnesses and tools. The project is open source and not tied to a specific model, broadening integration options. It can be self-hosted to operate privately on internal infrastructure, providing organizations with deployment and configuration options tailored to their constraints.