Buzz
An open-source AI agent collaboration workspace produced by Block, built on the Nostr protocol. It enables humans and agents to collaborate as equals through signed events. The workspace integrates multiple agents including Goose, Codex, and Claude Code, and supports opening pull requests, running code, and managing releases within channels.
Timeline (last 90 days)
All six types of agents — OpenClaw, Hermes Agent, Agent Zero, OpenHands, Paperclip, and Buzz — are fully open source, can all be self-hosted, and are not locked into any model ecosystem.
Buzz comes from Block (Jack Dorsey's company) and is essentially a self-hosted Slack, providing infrastructure for collaboration between agents and between people and agents; it is not itself an agent.
In tools such as Claude Code and Buzz, Claude's runtime itself is responsible for executing the agent loop, and the application layer only provides the UI, permission control and orchestration.
Buzz bridges to Claude via ACP (Agent Client Protocol), and the actual agent executor remains the Claude runtime.
The six agents in this article are ranked from easiest to hardest in terms of ease of onboarding and configuration: OpenClaw, Hermes, Agent Zero, OpenHands, Paperclip, and Buzz.
The design that decouples the runtime framework from the underlying models allows switching to cheaper models for lightweight tasks and switching to top-tier models for core business logic.
When an Agent completes a task in the background after being assigned work but does not reply in the public channel, it is because the routing is blocked rather than a permissions issue.
Multi-agent collaboration should follow two iron rules: frontend contract-first development (strictly following interface documentation, no self-initiated mocking) and backend prohibition of silent fixes.
The all-AI command mode has a single-point bottleneck, where the leader model may forget to @ a specific Agent, causing the group chat to go silent.
The fully automated pipeline mode lacks human correction, and early-stage requirement deviations can be rapidly amplified, leading to catastrophic rework.
13 more timeline events
All Facts (20)
The open-source voice tool ecosystem includes projects such as Vosk (offline recognition supporting 20+ languages), Kaldi, and desktop applications Buzz and MacWhisper.
60%UnverifiedAll six types of agents — OpenClaw, Hermes Agent, Agent Zero, OpenHands, Paperclip, and Buzz — are fully open source, can all be self-hosted, and are not locked into any model ecosystem.
50%UnverifiedBuzz comes from Block (Jack Dorsey's company) and is essentially a self-hosted Slack, providing infrastructure for collaboration between agents and between people and agents; it is not itself an agent.
50%UnverifiedBuzz bridges to Claude via ACP (Agent Client Protocol), and the actual agent executor remains the Claude runtime.
50%UnverifiedIn tools such as Claude Code and Buzz, Claude's runtime itself is responsible for executing the agent loop, and the application layer only provides the UI, permission control and orchestration.
50%UnverifiedThe six agents in this article are ranked from easiest to hardest in terms of ease of onboarding and configuration: OpenClaw, Hermes, Agent Zero, OpenHands, Paperclip, and Buzz.
50%UnverifiedBuzz's model layer can connect to models such as Claude, Kimi, and DeepSeek.
50%UnverifiedThe design that decouples the runtime framework from the underlying models allows switching to cheaper models for lightweight tasks and switching to top-tier models for core business logic.
50%UnverifiedWhen an Agent completes a task in the background after being assigned work but does not reply in the public channel, it is because the routing is blocked rather than a permissions issue.
50%UnverifiedThe solution to an Agent not replying is to inject a highest-priority rule into the system prompt, instructing it to call a specific CLI code when outputting and to specify the current channel UUID and the event ID to reply to.
50%UnverifiedMulti-agent collaboration should follow two iron rules: frontend contract-first development (strictly following interface documentation, no self-initiated mocking) and backend prohibition of silent fixes.
50%UnverifiedThe all-AI command mode has a single-point bottleneck, where the leader model may forget to @ a specific Agent, causing the group chat to go silent.
50%UnverifiedThe fully automated pipeline mode lacks human correction, and early-stage requirement deviations can be rapidly amplified, leading to catastrophic rework.
50%UnverifiedThe recommended V3 human-AI collaborative workflow is divided into three phases: human confirmation points, JSON contract alignment, and independent QA (where Claude, not involved in writing the code, tests requirements one by one).
50%UnverifiedThe prompt whitespace technique advocates setting broad goals, providing specifications, and leaving gaps so that AI can autonomously decide on specific data structures and locking mechanisms.
50%UnverifiedMulti-agent collaborative reasoning involves hardware-level network latency, and early versions of open-source tools may have occasional bugs.
50%UnverifiedBuzz adopts a three-layer architecture: Layer 1 CLI command-line layer, Layer 2 Harness runtime framework, and Layer 3 model layer.
50%UnverifiedBuzz introduces a Channel concept similar to WeCom and Slack, allowing multiple AI Agents to share context and collaborate within the same group.
50%UnverifiedBuzz has multiple built-in Agents including Goose, Codex, and Claude Code, which can directly open PRs, run code, and perform releases within channels.
50%UnverifiedBuzz is designed based on the Nostr protocol, where every action by a human or Agent is an event signed with a key and dispatched by type.
50%