AWS Built an Inbox for Its Agents. Chat Was Never the Right Interface for Work That Takes an Hour.
Here's a quick test for any AI agent you're thinking of paying for. Give it a job that takes an hour, then close the tab.
If the job dies with the tab, you haven't got an agent. You've got a very clever text box that needs you sitting in front of it the whole time, like a kettle that only boils while you're watching it.
On September 10, AWS open-sourced a project that takes this test seriously. It's called Pizza Bot (named after Amazon's two-pizza teams, not the food, sadly), and it's an email-style inbox for AI agents that work in the background. It reportedly began inside Amazon as a side project called JoeBot. By the time it was rebuilt for public release, more than 2,000 Amazon employees had used earlier versions for meeting prep, email drafting, Slack summaries, CRM logging and research.
The software is interesting. The reasoning behind it is more interesting, and the team put it in six words: chat assumes both parties are present.
Chat assumes you're both in the room
Every chat interface rests on an assumption so basic nobody notices it. You say something, the other side answers, and you both wait for each other. That works beautifully for questions. It falls apart for work.
The Pizza Bot team names three places where it breaks: when a task takes minutes, when it needs someone's approval, and when it runs on a schedule. As they put it, give an agent a task actually worth delegating and you'll be waiting a while.
Think about how you hand work to a human colleague. You don't pull up a chair next to their desk and watch them type. You explain the job, go back to your own day, and they come back when it's done or when they're stuck. The communication tool that grew up around that pattern is email. It's asynchronous by design, it keeps a history, and a message can sit there for three hours until you're ready for it.
So the best interface for delegating to software turns out to be one of the oldest ones we have.
The three queues are the whole idea
Pizza Bot sorts everything into three views: All (the full thread history), Unread (finished work you haven't looked at yet) and Action (work that's paused, waiting for your approval or your answer).
That's roughly how a good manager already thinks about a team: what's been done, what's done but not yet reviewed, and what's blocked on me. Action is the queue that matters. It means that when the agent hits a decision it shouldn't make alone, its default is to stop and ask, and the question doesn't expire. Approvals pause durably, so you can answer from your phone hours later and the work picks up where it left off.
Tasks can start by hand, on a cron schedule or from a webhook, and scheduled work produces threads you can read, not logs you have to dig through. That sounds minor until you've lost a morning in log files trying to work out what an automation did at 3am.
The underrated button is "edit"
When Pizza Bot wants to run a tool that's configured to need approval, it shows you the exact values it plans to use. You get three options: approve, reject or edit.
Most approval flows only offer the first two, which creates a quietly miserable pattern. The agent drafts an email to a client, it's 90% right, and your only choices are to send something slightly wrong or throw the whole thing away and explain the task again. Edit is how you'd treat a colleague: fix the one wrong figure, send it, move on.
Memory gets the same treatment. It's opt-in, it lasts across threads, and you can open it and change it. An agent that remembers things you can't see is a liability. An agent whose memory you can read and correct is closer to a colleague who takes decent notes.
The part everyone messes up: someone has to be home
One detail in the documentation deserves more attention than it's getting. If you run Pizza Bot as a desktop app and close it, the embedded server stops and any steps in flight may be lost. For work to carry on without you, you need an always-on backend.
That isn't a criticism. Pizza Bot is a community project with no AWS support or SLA behind it, and the team says plainly that keeping it running, backed up and up to date is your job. But it points straight at where the real difficulty lives. The inbox is only half the design. The other half is a worker that exists when you're not looking: a process that's still running at 3am, recovers when a connection drops, and doesn't need your laptop open to finish Tuesday's report.
When background agents fail, this is usually where it happens, not in the interface.
Where we'd push it one step further
Let's be upfront about our bias. This is the design argument we've been making for a while, so watching the biggest cloud company in the world arrive at the same place independently is rather pleasant. (We're trying not to be smug about it. We're failing slightly.)
But one difference matters. Pizza Bot's inbox is where your agents report to you. At Geta.Team, the inbox is also how everyone else reaches the agent. Each AI employee has its own real email address, so a client, a supplier or a partner at another company can write to it directly, the same way they'd write to a person. Work doesn't only flow from the agent to you. It also flows to the agent from outside, without you acting as the middleman who forwards everything.
That changes what the agent is. An inbox you check makes the agent a tool with good manners. An address other people can write to makes it a member of the team.
The rest of Pizza Bot's design we agree with completely: self-hosted, no telemetry, data in storage you control, your choice of model provider, approval before sensitive actions, memory you can edit. None of those are premium features any more. They're the baseline to expect before you let an agent anywhere near your email.
What to take from this if you're choosing an agent
You don't need to install Pizza Bot to use the lesson. Ask these five questions of any agent product:
- Does the work survive you closing the tab? If not, it's a chat assistant, whatever the landing page says.
- When it isn't sure, does it stop and ask, and does the question wait for you? Guessing is cheaper to build and far more expensive to live with.
- Can you correct an action without rejecting it? A good edit button saves more time than most benchmark gains.
- Can you read what it remembers? If you can't audit its memory, you can't trust its judgement.
- Can other people reach it directly? Or does every piece of work have to pass through you first?
Chat isn't going anywhere. It's still the fastest way to ask a quick question, and we use it all day. But once the work takes longer than the conversation, you want a colleague who goes away, does the job, and comes back with it finished or with a good question.
Want to test the most advanced AI employees? Try it here: https://Geta.Team
