Pizza Bot: Designing an Inbox-style Interface for Long-Running AI Agents
This article introduces AWS's open-source Pizza Bot, an application providing an email-style inbox for managing long-running AI agents. It addresses the architectural challenge of persistent agent-human interaction, especially when agents operate autonomously in the background and require human intervention or review at asynchronous intervals. The design emphasizes state persistence and asynchronous communication, diverging from traditional real-time chat interfaces for AI.
Read original on The New StackThe Challenge: Asynchronous Human-Agent Interaction
Traditional chat interfaces are ill-suited for AI agents that perform long-running, background tasks. The core problem is how to manage interactions where the human user might not be present, and the agent's work needs to persist beyond a single session. Pizza Bot tackles this by adopting an email-style inbox metaphor, treating agent updates and decisions as unread threads or actionable items that can be triaged at the user's convenience.
Core Design Principle
The fundamental design assumption is that "The interface assumes you are not watching." This shifts the burden of continuous monitoring from the user to the system, enabling pauses, notifications, and scheduled work to produce threads rather than just logs.
Architectural Components and Persistence
Pizza Bot's agent runtime is built on DeepAgents, which in turn utilizes LangGraph for stateful agent execution. A critical architectural decision here is the use of LangGraph's checkpointing mechanism. This allows an agent's state to be persisted throughout its workflow, enabling runs to pause for human approval, survive client disconnections, and resume later without loss of context. This persistence is crucial for supporting long-running, interruptible agent tasks.
- Agent Runtime: DeepAgents, built on LangGraph, handles the execution and state management of AI agents.
- State Persistence: LangGraph checkpoints agent state, allowing for pauses, disconnections, and later resumption.
- Data Storage: Checkpoints, threads, and other application data are stored locally using SQLite and ordinary files, supporting self-hosted deployment.
- Extensibility: Support for various LLM providers (Anthropic, Bedrock, Gemini, OpenAI, etc.) and extensibility via MCP servers and Agent Skills (e.g., Playwright for browser automation).
The server component can run locally, on an always-on host, or in a container, allowing scheduled agents to continue working even when the user's local machine is offline. This design enables the user to pick up agent threads from different devices, reinforcing the asynchronous, inbox-driven model.