Background agents need an inbox, not chat windows

Give an agent a task actually worth delegating and you will find yourself babysitting — watching a chat window scroll while it stops halfway, waiting for one approval. This week AWS open-sourced Pizza Bot (2026-09-10, Apache 2.0), turning “what background agents produce” into an inbox: finished work lands in Unread, work waiting for approval in Action. It looks like a small tool; it is actually a supervision pattern validated inside a big company: asynchronous agents need an asynchronous supervision surface. Supervision is queueing decisions, not watching logs.

Evidence: an agent interface built on outbox thinking

The AWS blog says it best: “You don’t send an email and then sit watching the outbox until the reply lands.” Pizza Bot is shaped like an email client for exactly that reason: a thread is a unit of work you come back to, not a session you have to attend. It began inside Amazon, where earlier versions were used by more than 2,000 people for meeting preparation, email drafting, Slack summaries, CRM logging, day prioritization, and web research, then rebuilt from the ground up as an open source project (AWS blog; sources at the end).

Architecturally it insists on “running where you can own it”: self-hosted, no telemetry; the server owns agent state and answers over HTTP, with desktop, browser, and terminal clients; runs are checkpointed, so quitting mid-run loses only the step in flight, not the whole thread. Model providers are your choice (Anthropic, Bedrock, Gemini, OpenAI, OpenRouter, even local Ollama); tools plug in via MCP; skills use the SKILL.md Markdown convention.

The two most valuable mechanisms are engineering-level: interruptOn is a per-tool approval policy — which tool calls must stop and wait for a human, with allowedDecisions setting the buttons you see; edit lets you correct an agent’s proposal instead of rejecting and rerunning from scratch. Installing an MCP server or plugin is an explicit decision — you confirm permissions before code runs.

Our judgment: asynchronous agents need asynchronous supervision surfaces

Chat assumes both parties are present: it holds for quick exchanges and breaks the moment a task runs for minutes or waits on your approval. The supervision surface for background and long-running agents must be “things you come back to,” not “processes you watch in real time.” Three design points are worth copying anywhere: first, separate output from pending decisions — queue “what the agent finished” (Unread) apart from “what the agent needs you to decide” (Action), so a returning human works by priority instead of digging through logs; second, make approval editable — correcting an agent’s proposal is cheaper than rerunning and closer to real collaboration; third, do not outsource supervision data — self-hosted, no telemetry, your own model, so output and approval records stay where you own them.

This is not theory for us. OOMeta’s daily cron pipeline is the same pattern: agents assemble signals into digest queues overnight, a human reviews and decides; batches with no change emit [SILENT] instead of noise. Unread/Action separation and our signal discipline are one principle — queue what needs a human decision, out of the noise. The larger your background-agent fleet, the more expensive this principle gets: every unsupervised asynchronous agent adds one more log nobody reads.

Implementation checklist: moving the inbox pattern into your background agents

Action: a two-week pilot that measures supervision cost

Pick one weekly workflow (nightly reconciliation, scheduled research briefs, batch document cleanup) and run it for two weeks with Pizza Bot or an equivalent inbox pattern, recording three numbers: how often human intervention was needed, how long each intervention decision took, and the share of agent output adopted without changes. The first number tells you how much automation is safe; the second tells you whether supervision itself is expensive; the third tells you whether agent quality has crossed the line.

The question for buyers: are your background agents “sending letters” or “rolling logs”? If the answer is the latter, the problem is not the agent — it is the supervision surface.

References: AWS Open Source Blog, “Introducing Pizza Bot, an open source inbox for AI agents that work in the background” (2026-09-10) https://aws.amazon.com/blogs/opensource/introducing-pizza-bot-an-open-source-inbox-for-ai-agents-that-work-in-the-background/ · Pizza Bot GitHub repository https://github.com/pizza-bot-app/pizza-bot