Amazon Web Services has officially open-sourced Pizza Bot, a self-hosted application designed to manage asynchronous AI tasks. Originally an internal tool supporting over 2,000 Amazon employees with meeting prep, email drafting, and CRM logging, the public release rebuilds the platform as a flexible, developer-friendly project. It moves beyond simple chat interfaces by organizing agent outputs into an email-style inbox, separating completed work from pending decisions. This structure addresses a common pain point in agentic workflows: keeping track of long-running processes that don’t require immediate attention but still need human oversight.
The architecture relies on DeepAgents and LangGraph for stateful execution, ensuring that tasks persist even if the user disconnects. A Hono API server handles runtime execution and storage, while clients built on Electron or React communicate via HTTP and server-sent events. Crucially, the system uses checkpoints to retain thread state. If you close your desktop app, the embedded server stops, potentially losing the step in flight. However, an always-on backend allows work to continue uninterrupted. Missed cron intervals trigger a single catch-up run rather than replaying every missed event, preventing redundant processing after downtime.
Pizza Bot supports a wide range of model providers, including Amazon Bedrock, Anthropic, Google Gemini, OpenAI, and local models via Ollama. Security and control are managed through “skills” defined in SKILL.md files, which scope tool access and define worker instructions. The system integrates with MCP servers to expose external tools, allowing agents to perform file operations or run sandboxed JavaScript. Users can configure specific policies to require approval for certain actions, giving them the ability to edit arguments or reject proposed steps before execution. This granular control is vital for enterprise adoption, where unchecked autonomous actions pose significant risks.
By releasing this under the Apache 2.0 license, AWS invites developers to build custom skills and integrate their own toolsets. The move signals a shift toward more robust, persistent agent frameworks that prioritize reliability over novelty. While many AI tools focus on speed, Pizza Bot emphasizes continuity and auditability. It remains to be seen how the community will extend its capabilities, but the foundation for serious, long-running autonomous workflows is now firmly in place.
