Pizza Bot, an open-source application from developers working at AWS, moves long-running AI agents into an email-style inbox instead of a chat session that requires constant attention. Released under Apache 2.0, the self-hosted tool lets agents run scheduled or webhook-triggered work in the background and notify users when results are ready. The design matters for business because it treats each thread as work to return to, with human approval built in before consequential actions.
How the inbox for background agents works
The application uses a client-server architecture and organizes work into All, Unread, and Action queues, plus an Activity panel that shows tasks delegated to specialist agents and their progress. Agents can inspect multiple sources, delegate parts of a task to specialized workers, pause for human approval, and then complete the job. The authors are Joseph Dolivo, principal technologist at AWS, and Igor Fil, solutions architect at AWS, who describe the product as offering developer detail without unnecessary complexity.
The system supports multiple model providers, MCP servers, and Agent Skills, while agent state, threads, checkpoints, attachments, and logs remain in a user-controlled folder on the local machine. Data leaves the machine only for the configured model provider and explicitly enabled MCP servers, with credentials kept in the OS secret store. The authors warn that enabled MCP servers can act on behalf of the user, so each installation choice carries operational consequences and needs review.
The project began in April 2025 as a side project called JoeBot, built to automate repetitive CRM logging, and later expanded after interest from non-technical colleagues. Its name refers to Amazon's two-pizza teams, a point Dolivo noted on LinkedIn. On Hacker News, the authors cited inspiration from a LangChain article about ambient agents, while early commenters praised the asynchronous and local-first model but questioned how many similar tools now exist and whether an open project can reproduce the skills and integrations of the internal version.
What this means for companies adopting agents
For companies, the inbox model changes how agent work fits into daily operations, since staff do not need to monitor execution and can process completed threads, approvals, and follow-ups in batches. Small teams gain a way to run scheduled research, data collection, or back-office tasks without dedicated infrastructure, while larger firms can test delegation across specialist agents before committing to enterprise platforms. The Apache 2.0 license and local storage lower the barrier to pilots that keep working data under company control.
The trade-off is that Pizza Bot is a community project rather than an AWS service, with no AWS support or service-level agreement. Teams remain responsible for hosting, backups, updates, and access control, and must verify which model provider and MCP servers are enabled for each deployment. The news by itself does not mean proven security, reliability, or parity with commercial agent platforms, so buyers should ask vendors and internal owners about logging, approval policies, and recovery from checkpoints.
A practical marker will be whether independent deployments, integrations, and Agent Skills accumulate outside the original team during the next few months. Steady contributions and documented production use would confirm demand for local-first agent inboxes, while stagnation would suggest companies prefer managed services despite weaker data control.
