Questions, inbox & @Agent
An agent working for you sometimes reaches a decision that belongs to you. It asks you in Everythings and stops. The question waits on the thing it is about, in the workspace's 📥 Inbox and in a notification, and you answer when you get to it: on the web, on your phone, or by telling the agent in the chat.
This page also covers the other direction, leaving a job for an agent with @Agent in a comment, and the labels that show what an agent wrote.
How an agent asks
An agent asks with the request_input tool. It is told to use it when a choice is yours to make (which of two approaches, whether to go ahead, what to call something) and you are not in the chat to be asked.
- The question goes on the thing it is about, so it reaches you with its context. When no thing fits, the agent creates a new thing to carry it.
- It can offer up to six options, each with a short label and an optional line of description. Six is what fits on a phone screen as buttons.
- With no options it is an open question, and you answer in your own words. A question can be up to 500 characters long.
- Having asked, the agent ends its turn. It is told never to guess and carry on, and never to answer its own question.
- A thing holds one open question at a time. An agent that tries to ask a second one on the same thing is refused until the first is answered or dismissed.
Asking is a write, so it needs a read & write connection. A read-only agent can still read the answers to questions asked before.
Where a question shows
One question shows up in several places at once, so you meet it wherever you are.
On the thing
A panel sits right under the thing's content, above its sub-things. It reads “✦ Claude Code is waiting on you” (with the name of the agent that asked), then the question, the options and a field for your answer. The panel hides while you edit the thing.
On its tile
In the grid, a ❓ sits in the bottom-left corner of the thing's icon while the question is open. Hover it to see which agent asked.
In the 📥 Inbox
While at least one question in a workspace waits, a 📥 Inbox tile appears beside Faves on the workspace's grid, with a badge counting the questions. It opens the Inbox, which lists them under Questions, each with the thing's name above it and an Open link beside the buttons for when you want to read the whole thing first. Closing the thing brings you back to the Inbox.
You answer right there. A question leaves the Inbox once it is answered or dismissed, and when the last one goes, the tile goes too. On the phone the tile opens the same questions in a sheet.
In a notification
The app notifies you on your phone and in the bell on the web. The title reads “Claude Code asks:” followed by the start of the question, with the thing's name under it. In a shared workspace every member is notified, each according to their own My agent asks me a question switch.
In the chat and in Claude Code
The view in the chat, the claude.ai panel and the Claude Code pane all show the question on its thing, read-only (Answer in the chat, below, says why). In Claude Code with the plugin, a band above the prompt counts your ❓ open questions across every workspace.
Answer in the app
The question panel works the same on the web and on the phone.
- Open the question from the notification, the 📥 Inbox tile, or the thing itself.
- Pick an option. The one you pick is highlighted; pick it again to clear it.
- Write anything you want to add in Add a note (optional). On an open question the field reads Your answer.
- Press Answer. With no option picked and only text written, the button reads Reply, since a written reply may well be a question back.
Dismiss closes a question that no longer matters, with no answer.
Once you answer, the panel shows who chose what, the date, and where the answer came from: on the web, on the phone, or answered in a chat. Your note shows under it. Your answer is also posted as a reply to the question in the thing's comments, so the whole exchange reads as one thread there too.
In a shared workspace any member can answer, and the first answer counts. If someone got there before you, the app says Already answered and shows their choice.
Answer in the chat
If you are talking to the agent anyway, just tell it. The agent records what you said with answer_request: the option you chose, your words, or that you declined. The thing then shows your answer as answered in a chat.
The agent is told to use answer_request only to relay what you actually said, never to decide for you. If you already answered in the app, its call changes nothing and hands it your answer instead.
Why the views in the chat are read-only
The Everythings view inside Claude, the claude.ai panel and the Claude Code pane all reach Everythings through the agent's own connection. A button there would press through that connection, so your answer would be recorded as the agent's, and nobody could tell it from the agent answering its own question. They show the question and, once you have answered, your answer. To answer, use the app, or say it in the chat.
Follow-ups and resolving
A question is a thread. It can take a few rounds, and the agent closes it when it has what it needs.
- When your answer needs a follow-up, the agent asks again on the same thing. The new question continues the thread as the next round, and you are notified again.
- The panel shows the earlier rounds above the current one, each with who said it and when, so a follow-up reads with what it follows up. The same exchange appears in the thing's comments.
- Between your answer and the agent's next move, the thread waits on the agent.
list_requestsmarks itwaitingOn: 'agent'until the agent follows up or resolves it, so an answer you gave stays on its list. - When it has what it needed, the agent closes the thread with
resolve_request. The panel then reads Resolved by and the agent's name. Resolving notifies nobody, and only an answered question can be resolved.
How the agent hears back
How soon an agent sees your answer depends on where it runs.
An agent in a chat you are in
It sees your answer when you tell it, or when it next reads the thing.
An agent that runs on its own
Every run starts with what_changed, which returns what people did since the agent's last run: answers to its questions, things they created, edited or trashed, and their marks and comments on its work. It then calls list_requests for the answered questions and reads the thing, its comments and its marks before carrying on. A scheduled agent therefore picks your answer up at the start of its next run.
Claude Code with the plugin
A session that asked you something carries on by itself. While it stays open, the plugin checks for your answer every five minutes, for up to two hours after the question, and wakes the session once you answer. Each question wakes it once. See Claude Code plugin.
Before it asks again
Each question an agent lists carries seenSinceAsked: whether anyone opened the thing on the web or the phone since the question was asked. When nobody has, the agent is told to wait, or say so in the chat, rather than ask the same thing again.
Jobs for @Agent
Questions go from the agent to you. @Agent goes the other way: you leave a job in a comment for whichever agent runs next.
- In a comment on the thing the job is about, type @ and pick Agent from the list. It is offered on every account, next to the workspace's members.
- Write the job, for example “@Agent please break this thing up into sub-things”, and post the comment.
Nobody is notified. The next agent run calls list_mentions right after what_changed, which lists every open @Agent job it can reach, oldest first. The agent reads the thing and your comment and does the work. Then it closes the job with resolve_mention: @Agent in your comment turns into 🤖agent, and a reply from the agent saying what it did joins your comment's thread. You are notified of that reply as of any reply to your comment. Each job is picked up once.
- When the job is unclear, the agent asks you with
request_inputon that thing and leaves the job open until it is done. - In a shared workspace, the job goes to the first agent run that reaches it, whichever member connected that agent.
- In Claude Code with the plugin, the band above the prompt shows 🤖 @Agent jobs with a count. Press it to list them in the pane.
Mentioning a person works the same way and notifies them. See Comments & mentions.
Labels on agent work
Everything an agent writes carries its name, taken from its connection: the app name it registered when you signed it in, or the name you gave its access token. The ✦ means “an agent wrote this”.
| Where | What you see |
|---|---|
| Thing page | A chip at the bottom right reads ✦ and the agent's name, such as ✦ Claude Code. Hover it to see whether the agent created the thing or last changed it. |
| Grid tile | A ✦ in the bottom-left corner of the icon. While the agent waits on a question there, a ❓ takes its place. |
| Comment | The byline reads “via” the agent, such as By: Sam · via Claude Code. |
| Mark | Hover the mark for the list of who left it: a mark your agent left shows your name with (agent) after it. |
| Notifications | Other members of a shared workspace get the usual notice for a new or edited thing, with “via” and the agent's name in the title. |
When you edit the thing yourself, the agent stops being named as its last editor. A thing the agent created still names it as the creator.
Stance and claims
Two more things an agent can put on a thing: how sure it is about what it wrote, and a claim that it is working on it.
How sure the agent is
An agent can attach a stance to a thing or a comment it writes: guess for an inference, likely for a partial check, or verified only when it ran or observed the thing itself, with an optional basis such as “ran the test”. Everythings records which agent said it and when.
Your ✅ done mark, left in the app after the agent wrote, counts as your verification: the next agent that reads the thing with get_thing sees who verified it and when. Agents are told never to set verified on your behalf. When you rewrite the text yourself, the stance is cleared, since it was about words that are gone.
Claims
When several agent runs pick from the same list, a run claims a thing with claim_thing before working on it and releases it with release_thing when it is done. Another run that tries to claim it is told the thing is held, and picks something else.
On the web, the thing page shows a chip with 🔒, the name of the machine the run is on and how long ago it claimed, such as 🔒 office-laptop · 12m. Hover it for the agent, its note and when the claim ends. In the grid, a 🔒 joins the marks over the icon.
The agent chooses how long its claim lasts when it takes it: four hours unless it asks for another length, and never more than 24 hours. Claiming again from the same machine renews it. A claim whose time has run out counts as none: the chip dims and reads expired, and another run may take the thing over.
If a run stopped without releasing its claim, use Clear claim in the thing's ⋮ menu. Any member can clear an expired claim, and the workspace's owner and admins can clear a live one too. A claim changes nothing else about the thing: whoever may edit it still can, and claiming notifies nobody.
Notifications
Two notification kinds are about your own agent. Both are on by default, and you choose them per workspace.
| Switch | What it tells you |
|---|---|
| My agent asks me a question | An agent asked a question on a thing in this workspace, including a follow-up round. |
| My agent creates or edits a thing | Your agent created, edited or deleted a thing in this workspace. |
You find them in the workspace's settings, on the Notifications tab: open the workspace switcher and press the gear on the current workspace's row. Each change saves at once.
Why these two exist
Everythings never notifies you of your own actions, and an agent acts as you: it writes with your account. Without these two kinds you would never hear about your own agent. With them, a workspace of one with a connected agent still tells you when the agent asks something or changes a thing.
On the phone, the app asks for permission to send notifications the first time you have a shared workspace or a connected agent.
Every other switch, the bell, push and Slack messages are on Notifications. To set up the routine that asks these questions, read The Agent workflow.
Common questions
Can I answer from my phone?
Yes. Tap the notification, or open the workspace and tap the 📥 Inbox tile. The phone shows the same panel as the web: the options, a note field, Answer or Reply, and Dismiss. Your answer reads “on the phone” on the thing afterwards.
Why can't I answer inside Claude's view or the Claude Code pane?
Those surfaces reach Everythings through the agent's own connection. An answer pressed there would be recorded as the agent's, and the record could no longer tell your choice from the agent answering its own question.
So they show the question read-only. Answer in the app, or tell the agent in the chat and it records your words with answer_request.
Someone else in the workspace answered first. What happens?
The first answer counts. In a shared workspace every member sees the question and any of them can answer it. If someone got there before you, the app says Already answered and shows what they chose and when.
I answered. Why hasn't the agent carried on?
It depends on where the agent runs. An agent in a chat you are in learns your answer when you tell it there. A scheduled agent picks it up at the start of its next run.
A Claude Code session with the Everythings plugin wakes on its own, within about five minutes, as long as the session is still open and the question is less than two hours old. See Claude Code plugin.
How do I get rid of a question that no longer matters?
Press Dismiss on it, on the thing or in the Inbox. The question closes without an answer, leaves the Inbox, and the agent sees that you dismissed it.
Who picks up an @Agent job in a shared workspace?
The first agent run that reaches it. list_mentions lists the open @Agent jobs in every workspace the agent's account can open, so an agent connected by any member can take the job. Once it is resolved, it no longer lists, so each job is done once.
Can I stop the notifications about my agent?
Yes, per workspace. Open the workspace's settings, go to Notifications, and untick My agent asks me a question or My agent creates or edits a thing. Each change saves at once. See Notifications.