Agent windows explained

An agent window is a named, living pane — Alex · Claude Code — that runs a real agent CLI on your machine and streams its work into a transcript. Around that transcript sit the controls you need to supervise: an input with Send, a Stop button while running, a token meter, a status glow, transcript Copy/Export/Clear, and a Loop this action that turns the last exchange into a self-improving loop.

Anatomy

• Title bar: name (rename anytime: rename Alex to Atlas), brand label, status dot, per-pane gear (styling), minimize/close.
• Transcript: your messages and the agent’s streamed output. Claude Code replies render live as they’re written — the reply text grows in place and each tool the agent runs shows as an activity line (⚙ Bash — npm test), so long tasks never look hung. Long sessions stay fast (old lines beyond 3,000 drop off). In Streaming Mode, secret-shaped strings render masked.
• Input bar: type and Send — or route from the command bar without touching the window.
• Token meter: running usage estimate for the pane.
• Glow & status: the border glows while running, with a brighter arc sweeping clockwise around it as an indeterminate progress indicator; the arc clears and a toast announces completion.
• Card flip: flip Alex shows the card’s back — a headline of what that agent was last asked to do, sized as large as the window allows, for at-a-glance supervision. It arrives in two stages. Instantly, on your machine, your prompt is shortened: politeness, lead-ins and “and reply in X” delivery wording come off and the actual ask stays. Then, if a local agent CLI is available, that headline is replaced by a true one-line paraphrase — the same meaning re-worded, so a rambling instruction reads as a title instead of a shorter ramble. Hover the quote for the prompt in full. The card never waits: the instant version needs nothing but your machine and is what you keep offline, and only the paraphrase step passes your prompt text to a model (whichever agent CLI is installed — Claude Code by default). Each distinct prompt is paraphrased once and remembered, so flipping the same card again is immediate.

The profile behind the name

Every named agent has a persistent, on-disk profile: work history (categorized automatically), skills, a backlog, and notes. It accumulates across sessions and windows — the identity is the name. This is also what Hermes consults when routing work.

Headless workers — agents without a window

An agent doesn’t have to occupy screen space. A headless worker runs the same way a window does but has no window of its own, so how much work can run at once stops being a question of how many windows fit on your canvas. Graphs, Super, Crew, Kanban assignment and Hermes can all send work to one.

Add and name them in Settings → Agents, where you also set how many may run at once. Identity works exactly as it does for windows — it’s the name. A headless worker called Scout shares the profile above with a window called Scout: same notes, same history, same memory.

Two deliberate limits. A headless run is a stateless one-shot, so Claude keeps using a window when you want its persistent process — a cold spawn each time would be slower, not faster. And there is no account failover without a window: an authentication failure fails that step honestly rather than quietly moving to another account.

Under the hood

Each message runs the agent’s CLI non-interactively with your local credentials. The agent also receives a one-line identity preamble (who it is in your workspace), so addressing it by name always makes sense to it.