Agent sandboxes

Gibson runs each dispatched agent, and each tool it calls, in its own ephemeral sandbox. The Agent Sandboxes page shows those runs while they happen. Open it at /dashboard/sandboxes.

The page is a wall of tiles. One tile is one run. Each tile streams that run's output as the agent produces it.

What the page shows

The wall lists everything your tenant is running in a sandbox right now. This includes the coding agent, but it is not limited to it. Any agent a mission dispatches, and any tool it runs, appears here.

On a tile	What it means
Status dot	Green and pulsing while the run streams. Grey when it finished or stopped. Red on a stream error.
Agent name	The name of the dispatched agent.
Run id	The first eight characters of the run id. Hover for the full id.
Kind	Whether this run is an agent or a tool.
Sandbox class	The setec isolation class the run was launched under.
Turns	Agent turns so far in this run.
Elapsed	Time since dispatch. It stops when the run stops.
Cost	Spend on this run so far, in US dollars.

Tools appear here too. Gibson runs a tool in its own sandbox, the same way it runs an agent. A port scan and a coding agent sit side by side on the wall. The kind chip on each tile says which one you are looking at.

What the page cannot do

The surface is read-only. It renders events and nothing else. There is no input box, no terminal session, and no path back to the agent. To drive an agent yourself, run it from your own machine.

Your tenant is the only scope. The server derives the tenant from your identity, so a run that belongs to another tenant is invisible and its run id is not resolvable.

How the wall fits your runs

The grid adapts to the number of running agents.

Running agents	Columns
1	1
2 to 4	2
5 to 9	3
10 to 25	5
More than 25	6

Two controls sit in the header. Density switches between comfortable and compact tiles. Sort orders the wall by start time, by name, or by cost. Both choices persist in your browser.

Live streams and the queue

A browser cannot hold many server-sent streams at once, so the page caps them. Only tiles you can see hold a live stream. A tile you scroll away from releases its slot, and the next tile in the queue takes it.

The header reports this as live N/cap, and adds M waiting when tiles are queued. A tile you scroll back to resumes from the last event it saw. The tail fills in without a gap and without a repeated line.

Full screen

Click a tile, or press Enter or F on it, to open that run near full screen. The pop-out shares the tile's stream, so nothing reconnects and nothing replays.

Key	What it does
Esc	Close the pop-out and return focus to the tile.
Left	Go to the previous running agent.
Right	Go to the next running agent.

A rail beside the output shows the run's facts. These are status, agent, run id, sandbox, sandbox class, model, session, start time, elapsed, turns, cost and agent time. The rail also offers Open mission, Stop mission and Copy run id.

The address bar carries the open run as ?run=<id>. That link is shareable inside your tenant. It opens the pop-out for that run.

After a run ends

A finished run stays on the wall for one minute with a ribbon that reads Completed, Failed or Stopped. It then folds into the Recent runs list under the wall, with a link to its mission.

Getting there

The sidebar entry is Sandboxes. It carries a live count of the agents your tenant is running. The mission page, the missions list and the agents page each show a Live sandbox link. That link appears while the mission or the agent has a run in flight.