Pizza Bot Approval Fatigue: The Hidden Cost of Agent Inboxes

An inbox instead of a chat window

AWS has open-sourced Pizza Bot, a self-hosted application that changes how people supervise AI agents. Instead of a conventional chat window, users get an inbox. Each delegated task lives in its own thread, and a queue collects work that is finished or waiting on a human. 2 AWS argues that background agents don't always need a user's attention while they run. An inbox lets people hand off long tasks, come back later, and see at a glance what is done and what needs intervention. 2

The tool is Apache 2.0-licensed and uses a client-server architecture. Agents can run scheduled or webhook-triggered jobs, delegate to specialized worker agents, and pause for human approval when needed. 3 It supports multiple model providers, MCP servers, and Agent Skills, and it keeps data and agent state on the user's own machine. 3

AWS principal technologist Joseph Dolivo and solutions architect Igor Fil explain the design with an email analogy. Nobody sends a message and then stares at the outbox until a reply arrives. In Pizza Bot, a thread is a unit of work you return to, not a session you have to babysit. 3

Not an isolated idea

Pizza Bot is part of a broader move toward asynchronous agent work. Anthropic's Claude Code offers a similar pattern through its agent view, which is still labelled a research preview. From one screen, a developer can launch and monitor several full Claude Code sessions, none of which needs an attached terminal. 1 Sessions can be started from a dashboard, launched from the shell with a background flag, pushed into the background mid-session, or forked so a copy keeps working while the original stays attached. 1

Both designs make the same choice: background work stops when it hits a permission boundary. In Claude Code, background sessions use the same permission model as interactive ones. A session waiting on a permission, sandbox, or MCP prompt moves to a "Needs input" list and waits there instead of proceeding silently. Fully unattended operation requires explicitly opting in to a bypass. 1 Pizza Bot's approval pauses and its queue of items needing human input reflect the same philosophy. 23

Where the hidden cost shows up

This is where the inbox metaphor has a weakness. Email works because most messages can be skimmed, archived, or ignored. Agent approval requests can't be handled that way, at least not safely. Every item in the queue is a decision. When agents run in parallel, the number of decisions grows with them.

Claude Code's documentation notes that running N background sessions consumes roughly N times the quota of a single session. 1 Throughput can scale with how many agents you launch. Human review capacity does not. Research on code review cited alongside these tools suggests a reviewer's sweet spot is about 200 to 400 lines of code per sitting. Google's engineering guidance treats around 1,000 lines in a single changelist as usually too large. Fatigue tends to set in after about an hour of continuous review. 1

Those figures come from human code review, not agent approvals, so applying them here is an inference. The direction still seems clear. An inbox that fills with completed diffs and permission prompts faster than a person can carefully evaluate them creates pressure to rubber-stamp. Approval fatigue could then weaken the very safeguard that blocking-by-default is meant to provide. The reviewer becomes the bottleneck, and the easy fix, broad auto-approval, gives back the control these designs were built to protect.

Governance and integration remain open questions

Industry analysts quoted on Pizza Bot's launch pointed to a related set of concerns. Integration and governance requirements, they said, could slow enterprise adoption. 2 A self-hosted tool that keeps state on the user's machine is good for privacy and control. 3 Organizations will still need to know who approved what, under which policies, and how agent actions connect to existing identity, audit, and workflow systems. An inbox makes pending decisions visible. It does not, by itself, make them accountable.

The takeaway

The move from chat to inbox is a sensible response to agents that run for minutes or hours rather than seconds. Both AWS and Anthropic have also made the responsible default choice: pause and ask rather than proceed silently. 12 But these tools change the job of the human in the loop. Users stop typing prompts and start triaging approvals.

The practical lesson for teams is to size agent work so it can actually be reviewed. That means keeping changes small, batching approvals sensibly, and treating the human review budget as a real constraint alongside model quota. Interfaces like Pizza Bot make parallel delegation easy. Whether they stay safe at scale will depend on how well teams manage the queue that piles up behind them.