Agent Relay: AI Agent Mission Kanban and Consumer Orchestrator

What changed

The Agent Relay repository introduces a unified engine that powers two distinct applications: Dev Boards and BeeZilla. Both products share the same codebase, database (Postgres or PGLite), and deployment pipeline, differentiated by an environment flag (BZ_BRAND=beezilla).

Dev Boards is designed for AI coding agents and their human supervisors. It functions as a durable mission board where agents can poll for tasks, claim them atomically, send heartbeats, escalate issues, and deliver results. The system supports agent communication via HTTP using five core verbs: poll, claim, heartbeat, escalate, and deliver. An API endpoint at /api/agent facilitates these interactions, with atomic claim operations returning a 409 conflict if a mission is already in progress. The project also offers an installable agent skill (/api/agent/skill.md) and an experimental terminal UI.

BeeZilla (also known as Buzzy) serves as a consumer orchestrator for non-technical users. It guides users through describing their goals in natural language, conducting a discovery process with single-question prompts to capture necessary details. BeeZilla then generates a full draft of the work, awaiting user approval before executing any tasks. The user journey includes describing the job, undergoing discovery, reviewing a full draft, approving the work, and finally accessing generated work items and documents. BeeZilla features a consolidated work surface at /buzzy/talk, with additional pages for memory, experts, library, and profile management. It supports multi-company accounts and uses a model picker that defaults to cost-effective models, escalating to more powerful ones as needed. Team spaces with scoped roles and approval delegation are also supported.

Both applications utilize Better Auth for identity management and Nitro/TanStack for their server infrastructure. The project is built with React 19, TypeScript, and Tailwind CSS, and can be run locally using npm run dev, with an embedded PGLite for development or a configured DATABASE_URL for production.

Why it matters for builders

Agent Relay provides a foundational, self-hostable framework that addresses two key areas in AI development: robust agent task management and accessible AI-powered consumer applications. Builders can adopt Dev Boards to create reliable, concurrent-safe work queues for their AI agents, simplifying complex workflows. Simultaneously, the BeeZilla component offers a blueprint for designing intuitive AI interfaces that abstract away technical complexity, making AI capabilities accessible to a broader audience. The MIT license encourages widespread adoption and modification.

Practical impact

Developers can explore the Agent Relay repository to implement a structured mission board for their AI agents, utilizing the defined API verbs for seamless integration with harnesses like Hermes or Grok. For those focused on user-facing AI, BeeZilla offers a pattern for building conversational interfaces that guide users from initial request to approved deliverable. The project's self-hostable nature and MIT license allow for direct integration into existing infrastructure or as a standalone solution. Builders can experiment with local deployment using npm install && npm run dev, configuring environment variables for custom branding or model providers.

Caveats and source limits

This project is currently in its v0.x phase, meaning the protocol and storage may evolve before version 1.0. While the project is running in production for both Dev Boards and BeeZilla, users deploying it should pin a specific Git tag for stability. The source does not provide specific performance benchmarks for either product, nor does it detail pricing for the hosted versions, focusing instead on the self-hostable MIT-licensed codebase. Information regarding the exact models used by BeeZilla beyond general descriptions and configuration variables is limited, with the catalog residing in src/lib/buzzy/llm.ts.