sessions
In short. AgentBox runs Claude Code sessions itself and renders each one as a conversation: live streaming, thinking folded away, tool calls one line each. A hotkey rolls the same sessions down over your work. A question from an agent in a session is answered inside that conversation, not in a card over it.
Read on if you want the agent and the answer in the same window. Skip to Work it does on its own for the runs nobody starts.
The one case where a question does not get a window: it renders in the conversation it is about, and never takes the composer's keyboard.
A session is a real claude child, headless, driven over its event stream. What
you get is not a terminal: prose set to a reading measure, code highlighted with a
copy button, tables with columns, and tokens arriving as they are produced.
Every agent turn carries the time it arrived on a 24-hour clock and how long the model worked before its first word, so a slow turn is legible as slow rather than as broken.
The mechanics of a turn collapse so the conclusion gets the space. Thinking sits behind one caret you can open. A tool call is a single line with the tool's name, its argument highlighted when the argument is code, and a mark saying whether it finished, failed, or is still going. That shape was arrived at by having it the other way first: a reply that ran five commands showed five bordered pills stacked down the column, each shouting as loudly as the answer.
Enter sends, Shift + Enter gives you a newline, and Esc hands the keyboard back out of the composer.
Warning
Plan and Full are the two buttons at the top of a session, and Full is full: the child runs with its permission prompts bypassed, so it edits files and runs commands without stopping to ask you. Plan can look and cannot change anything. The mode is fixed when the process spawns, which is why switching starts a new child and carries the conversation across rather than changing its mind.
Sessions are a list with a dot on whichever ones are working, plus New. Each one names itself: the heading the agent opens with, or its first sentence, cut down to about 42 characters. Until it has said anything the prompt you typed stands in.
Asking the model for a title instead would be another round trip and another entry in the transcript, for a worse answer than the sentence it already wrote.
Rename them if you want. Each is saved to disk as its own file with a markdown sibling, so a conversation from Tuesday reopens on Friday and is searchable in between.
Ctrl + Alt + the backtick key drops a session console from the top edge of the monitor your pointer is on, over whatever you were doing, and the same key rolls it up. So does Esc. It shows the sessions the app window shows, which is the point: dropping it down mid-task lands you where you already were, rather than in a second place your work might be.
This is the one AgentBox window that takes the keyboard. Nothing else does, and the reason is the direction of the interruption: a card is AgentBox arriving at you and must never steal a keystroke, while the console is you summoning AgentBox and you are about to type in it.
It does not slide, and that was measured rather than assumed. The animation holds 38 frames in 222 milliseconds, about 170 a second, more than the display shows, and it still does not read as sliding: this window grows its own height, so every frame is a re-composite rather than a texture moving, and the eye reads a growing rectangle. The roll is off by default and there is a setting for whoever disagrees.
When the agent asking is one running in a session, the question does not get a window. It appears in that conversation, directly above the composer, wearing the severity bar a card would have worn.
The reason is what a card would cover. The thing the question is about is on screen right behind it: the file it read a moment ago, the command it ran, the plan it wants a word on. A window over the middle of the screen hides the one thing you needed to answer it.
It still never takes the keyboard, which is the rule a card is built around and this surface inherits. The composer keeps focus, the single-key answers work only while you are not typing, and the hint line says which of those two you are currently in.
For a while there was a hole in this, and it was the worst possible hole. A host meant the app window only, so a question asked while the drop-down console was down opened a card over the very conversation it belonged to. Since 2026-08-06 the console is a host too. With both open the question renders in both, and answering either one resolves it.
Losing the host is handled the other way round: close the window with a question still pending and it comes back as a card, because an unanswered question must not leave with the surface it was rendered on. It does not work in reverse, on purpose. Opening a session window does not pull a card that is already on your screen back inline, because a question you can see is not a question that needs moving.
Only three kinds go inline: a choice, a confirm, and a notice. Anything that needs a text field, a secret, a form, a diff or a countdown gets the full card, and so does anything at urgent level whatever kind it is. A masked credential field belongs on a surface built for it, and urgent means the screen, not the window you might have closed.
A session is Claude Code specifically. It is spawned as a headless child with AgentBox's own tools wired in, which is what lets a session ask you something in the first place, and it is why a run of an assignment can be read here like any other conversation.
One layout cost worth knowing, found by looking at a real long question rather than at the diff. At a window 1000 pixels by 520, a long question left the reply box a clipped sliver with no Send button, so it could be read and not answered. The panel yields in two stages now: its body absorbs the squeeze first, then the whole panel scrolls as a unit, and the composer never shrinks at all.
Next: the work it starts without you, or everything a question can look like.