AWS Introduces Pizza Bot: An Open Source Inbox for Background AI Agents

Details

Amazon Web Services has open-sourced Pizza Bot, an asynchronous task inbox for managing background AI agent workflows. It solves a concrete problem in agentic systems: where do results land when AI tasks finish while the user is working on something else? The app organizes completed work into an email-style inbox with three views — All (full thread history), Unread (completed tasks awaiting review), and Action (tasks paused for human approval or input). The tool was battle-tested inside Amazon where earlier versions served over 2,000 employees, covering email drafting, meeting preparation, Slack summaries, CRM logging, and research tasks. The public release is a rebuilt open-source version under Apache 2.0, with macOS, Windows, and Linux desktop builds plus browser and terminal clients connecting to a local or standalone backend. Under the hood, Pizza Bot uses DeepAgents and LangGraph for stateful execution. A Hono API server handles runtime while Electron and browser clients share a React interface over HTTP and server-sent events. LangGraph checkpoints preserve thread state and approval pauses through disconnects; SQLite stores hold cross-thread memory and app metadata. One important constraint: quitting the desktop app kills its embedded server and any active runs — a persistent always-on backend is required for truly continuous background operation. All major AI providers are supported: Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter, and Ollama. MCP servers expose external tools to agents, and existing Claude Code-compatible .mcp.json configurations work without changes. SKILL.md files define worker instructions and scoped tool access. Skill authors configure approval requirements so users can accept, edit proposed arguments, or reject specific tool calls — giving genuine human-in-the-loop control.