AWS open-sources Pizza Bot, an email inbox for background AI agents
On September 10, AWS open-sourced Pizza Bot, an app that replaces chat with an email-style inbox for tracking AI agents working in the background. For any team running autonomous agents in the cloud, the thesis fits in one sentence: the interface must assume no one is watching.
September 10, 2026. AWS released Pizza Bot, an open-source application that gives AI agents an email-style inbox. Joseph Dolivo, principal technologist at AWS Startups, co-created it with Igor Fil. Two thousand people inside Amazon were already using early builds before the project was rebuilt as a standalone community effort. The bet is captured in a single line from the launch post: “The interface assumes you are not watching.” Nothing else starts from that assumption, and that assumption is precisely what changes everything.
Chat is the wrong tool for background agents
A background AI agent — one that watches a CRM, processes logs, or prepares a report overnight — does not look like a conversation. Chat assumes a human is present, initiating, prompting, and closing the loop. A background agent keeps working after the human has shut their laptop for lunch or for the night.
So Pizza Bot borrows the mechanics of email: finished jobs arrive as unread threads, anything needing a human decision is surfaced for action, and an Activity panel exposes jobs handed off to specialist agents — their tool use and progress. The metaphor is not cosmetic. It inverts the monitoring burden: the agent reports in, the human no longer chases it.
Dolivo’s framing sums up the reversal: instead of having to initiate every conversation or wait on a prompt, scheduled agents run autonomously and surface updates into the inbox for review and triage.
A community project, not an AWS service
Despite its Amazon roots, Pizza Bot is not an AWS service. The project lives in its own GitHub organization, with no Amazon support or SLA: it is entirely self-hosted.
It ships as a desktop app for macOS, Windows, and Linux, with browser and terminal clients available too. By default it starts a local Pizza Bot server on the machine, and the developer picks the model behind it: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local model via Ollama. The server can also run on an always-on host or in a container, so scheduled agents keep working with the laptop closed and their threads can be picked up later from another device.
Extensibility comes through MCP servers and Agent Skills. The bundled browser-automation skill uses Playwright MCP to navigate and interact with websites.
Under the hood: LangGraph and ambient agents
The agent runtime is built on DeepAgents, LangChain’s open-source harness for long-running agent tasks, which itself runs on LangGraph, its runtime for stateful agent execution. The decisive piece is persistence: LangGraph checkpoints an agent’s state as it works, letting a run stop for approval, survive a disconnected client, and resume later without starting over. Pizza Bot stores those checkpoints, along with threads and application data, in SQLite and ordinary files.
AWS already ships its own agent SDK, Strands Agents, available since May 2025 and framed as a lighter-weight alternative to LangGraph for agents that do not need explicit graph control flow. Asked directly, Dolivo concedes Strands could have worked, especially now that it supports TypeScript and workflows. But the team went with LangGraph “for the maturity of the tooling and breadth of the ecosystem,” and because it is more familiar to developers — a community-adoption argument.
Pizza Bot extends an idea LangChain introduced in January 2025, when Harrison Chase coined the term ambient agents: agents that respond to events, work concurrently, and involve a human only when needed. LangChain’s reference implementation was an email assistant built on LangGraph, fitted with an Agent Inbox, a standalone interface inspired by email and customer-support software.
From “JoeBot” to Pizza Bot
The origin story goes back to April 2025. Dolivo started what he calls a “side-of-desk passion project” named JoeBot, which automated repetitive CRM logging. He later teamed up with Igor Fil to turn that script into Pizza Bot, an MCP server that could execute parameterized, deterministic “recipes” against its internal systems.
The project quickly outgrew its engineers. It eventually reached more than thirty contributors and over two thousand users inside Amazon, and it needed a front end non-technical people could open. “An MCP server requires an MCP client, and expecting non-technical users to work out of an IDE or terminal was never going to cut it,” Dolivo says. The result is the version released this week: a desktop app built around an inbox rather than a terminal.
Why an inbox, why now
The timing is not a coincidence. OpenAI opened its Agents API on September 10, the same week Pizza Bot shipped: both announcements answer the same shift, agents moving from synchronous calls to durable background work. An assistant that answers a request and returns control is a product; an agent that watches, schedules, and re-runs is a process. The difference is one word: persistence, and persistence changes the interface.
The signals date back to 2025. Harrison Chase’s ambient agents already described agents that react to events rather than prompts, and LangChain built an Agent Inbox to prove it. What Pizza Bot adds is proof at the scale of an organization: two thousand users inside Amazon validated that an inbox is the missing abstraction.
Concretely, the work loop becomes: a scheduled agent wakes on an event, runs a recipe, checkpoints its state into SQLite, and drops a thread into the inbox. The human opens the app later, sees the unread threads, approves what needs approving, and closes it. Nothing required continuous presence — which is exactly the promise chat cannot keep.
What this changes for cloud teams
Pizza Bot is not a product you buy; it is a position on the interface for agents.
- Chat does not fit asynchronous work. An agent that runs for an hour during your lunch break does not produce a conversation, it produces state. An inbox is the interface that accepts that.
- Persistence is the real problem. The LangGraph checkpoint is what turns a fragile script into a resilient agent: a run can survive a disconnect and resume without starting over.
- Model neutrality matters. Plugging in Bedrock, Gemini, OpenAI, or Ollama interchangeably makes Pizza Bot a client, not a platform — which is exactly what makes community adoption possible.
The stakes outlive the tool. As agents shift from synchronous calls to durable background work, the monitoring interface becomes the product: who triages, who approves, who resumes after a crash. Pizza Bot answers by saying it is an inbox.
Verdict
Pizza Bot matters most as a prototype for a conviction: background agents need an inbox, not a console.
If you deploy agents that work unsupervised, try Pizza Bot as a triage-and-resume harness — the email mechanics and the LangGraph checkpoint solve the two real pains of asynchronous work: monitoring and resilience. If you want a managed service, look elsewhere: Pizza Bot is self-hosted, has no SLA, and replaces neither Strands nor an orchestration platform. The signal is not the product but the thesis AWS chose to carry: the interface that assumes no one is watching.