Evaluating Buzz: The AI Agent Group Chat

The Breakthrough

Buzz functions as a persistent, Slack-like group chat interface that allows multiple AI agents to coordinate, debate, and maintain a shared, searchable history of their actions and reasoning.

What Actually Worked

• Adversarial Review: Agents can be assigned opposing roles to critique project requirements, allowing for cross-model verification where one agent attacks a PRD while another defends it.
• Persistent Audit Trail: Every message is assigned a unique ID, creating a searchable record of which agent performed a task and which user initiated it.
• Compute Sharing: The platform allows users to chain multiple local machines together to host larger models, distributing the load across a cluster of hardware.
• Memory Export: Agent states, including their accumulated notes and project context, can be exported as JSON or PNG files to be imported into other environments.

Before / After

• Token Efficiency: A simple greeting response consumed 31,000 tokens in Buzz, compared to approximately 4,000 tokens for the same task performed via Claude Code in a terminal.

Context

Buzz is an open-source project from Block that attempts to move AI agent coordination out of the terminal and into a collaborative chat environment. While it solves the visibility problem of agent teams, the current implementation suffers from significant token bloat because it sends the entire conversation history to the agent even when the agent already maintains its own internal context. Furthermore, the lack of end-to-end encryption and the inability to restrict agent access to specific channels make it difficult to use for sensitive team environments.

Notable Quotes

• "Buzz is less a Slack alternative than a near perfect record of what your agents did, with almost no control over what they were allowed to do."
• "The same conversation is now in two places at once and you're using double the tokens you originally needed to use."