AWS Open-Sources Pizza Bot, an Email-Style Inbox for AI Agents

If you've ever left an AI agent chugging away on a task, then found yourself refreshing the chat window every thirty seconds like it's a pizza tracker, AWS has a fix. Its engineers have open-sourced Pizza Bot, a tool that reframes agent interaction as an inbox instead of a live conversation.

The pitch is simple and, honestly, kind of obvious once someone says it out loud. Chat interfaces assume both parties are sitting there, present and attentive, which works fine for a quick back-and-forth but falls apart the moment a task takes several minutes to complete. Pizza Bot's answer: treat agent output like email. Threads sit in an inbox. Finished work lands in an Unread folder. Anything that needs a decision from you gets flagged as Action. You check in when there's something worth checking in for, not before.

"You don't send an email and then sit watching the outbox until the reply lands," the Pizza Bot team wrote in the announcement post. "Pizza Bot is shaped like an email client for the same reason: a thread is a unit of work you come back to rather than a session you have to attend."

That framing extends to how the team talks about agents generally. "Pizza Bot treats an agent the way you'd treat a colleague who has gone off to do the work: it comes back when there's something to read, or something only you can decide," they added. It's a small reframe, but it matters if you're managing more than one agent at a time, which is increasingly the whole point of agentic workflows.

Where the design came from

The name is a nod to Amazon's "two-pizza team" model, small autonomous teams sized so two pizzas can feed the whole crew. Pizza Bot itself started life inside Amazon as an internal tool for non-coding agent work. According to the developers, every internal use case kept converging on the same shape: agents working autonomously in the background, only interrupting the user when a task finished or a decision was needed.

"None of those describe how anyone uses a chat window," the team explained. "They describe how people already work with each other." That observation pushed them toward email as the interaction model, and eventually toward open-sourcing the project to see if the pattern held up outside Amazon's walls.

Under the hood, Pizza Bot is a local-first desktop app, available for Windows, macOS, and Linux via its GitHub repo. A few implementation details worth flagging for anyone evaluating it:

• All communication threads are stored locally in SQLite files, and nothing leaves the device unless you explicitly grant permission.
• Prompts and attachments do get sent to whichever model provider you're using, as you'd expect, but Pizza Bot also supports Ollama for running open-weight models entirely locally if you want to skip that step altogether.
• Out of the box it's fairly light on capabilities: it can list, read, write, edit, and search files, delegate subtasks to specialist agents, and run a sandboxed JavaScript interpreter.
• It ships with a browser automation skill and a guide skill for asking Pizza Bot about its own capabilities.
• Claude Code-compatible .mcp.json files and skills can be dropped straight into the app, which should make it easy to extend for anyone already building in that ecosystem.

The maintainers were upfront that Amazon-specific skills and internal MCP servers were stripped out during the rebuild for public release, so what you get is closer to a clean slate than a fully loaded internal tool. That's probably the right call for an open-source release, but it does mean you'll need to bring your own extensions if you want it doing anything beyond the basics.

Worth flagging too: this is explicitly not an official AWS product. "There's no AWS support or service-level agreement behind it," the team notes. If you deploy this in a production workflow and something breaks at 2 a.m., you're on your own, same as with most open-source tooling, but worth saying out loud given the AWS branding.

The async-first approach also fits into a broader trend of agent tooling trying to get away from the single-session chat paradigm. As more teams run multiple agents concurrently, doing research, writing code, monitoring systems, the interaction model needs to scale beyond "stare at one window." An inbox, with its built-in affordances for triage, threading, and deferred attention, is a reasonable answer, and it's one that's culturally legible since basically everyone already knows how to use email.

Key Takeaways

Pizza Bot is a solid example of borrowing a decades-old UX pattern to solve a genuinely new problem. Chat interfaces were the default for AI agents mostly because that's how the first wave of LLM products shipped, not because synchronous presence is actually the right model for long-running autonomous tasks.

For engineering teams juggling multiple agents on background jobs, code review, web research, or anything else that takes more than a few seconds, the inbox metaphor offers a familiar mental model: triage what's urgent, batch what's not, and stop babysitting a spinner. The local-first storage and Ollama support are also worth noting for anyone with data residency concerns or a preference for running models on their own hardware.

Just don't expect AWS to pick up the phone if it breaks. This is a community project wearing an AWS logo, not a managed service, and the maintainers are refreshingly clear about that distinction.