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

Most agent demos assume you’re staring at the chat. Real work doesn’t. AWS’s Pizza Bot treats long-running AI tasks like mail: All for history, Unread for finished work waiting review, Action for jobs paused on approval or an answer.

Earlier internal versions served more than 2,000 people at Amazon for meeting prep, email drafting, Slack summaries, CRM logging, and research. The public rebuild is Apache 2.0, with macOS, Windows, and Linux desktop builds plus browser and terminal clients against a local or standalone backend.

We’re interested because the product metaphor is right.

Background agents fail socially before they fail technically: people lose track of what finished, what needs a yes, and what quietly went sideways. An inbox with Action items is how you make autonomy legible. Manual starts, cron, and webhooks cover the trigger surface; the server owns scheduling so downtime produces one catch-up run instead of a cron stampede.

Under the hood it’s DeepAgents and LangGraph for stateful runs, a Hono API, SQLite for memory and metadata, and SSE so reconnecting clients can replay buffered events. Closing a client doesn’t stop the server—but quitting the desktop app kills its embedded server and can lose the in-flight step even if checkpoints keep the thread. That’s an honest limitation: always-on backend required if you want work to survive walking away.

We’re for this pattern of useful AI—delegation with brakes. Skills via SKILL.md, MCP tools, sandboxed JS without network or host filesystem, and interruptOn / allowedDecisions so users can approve, edit, or reject tool calls. Multi-provider support (Bedrock, Anthropic, Gemini, OpenAI, OpenRouter, Ollama) keeps it from being a Bedrock-only toy.

The downside if inbox-agents scale carelessly: approval fatigue, overly broad folder grants, and cron jobs that keep touching CRM and mail while nobody’s watching Unread. Self-hosted helps; discipline still matters.

I’m watching whether outside Amazon the Action queue becomes a habit—or users disable interrupts and let the agent run hot.

MarkTechPost’s deployability note: one SQLite data directory supports one backend process. Plugins can package skills with MCP servers; Claude Code-compatible .mcp.json configs are supported.