Open Agents in the sidebar. The page carries one header with three sections under it, and it opens on the first: what needs a decision from you.

Moving between sections keeps you on the same page, so the browser Back button leaves the Agents home rather than walking you backwards through the sections you looked at. The New agent button sits in the header on every one of them. See Building Your Own Agent.

Two sentences can appear directly under the section control, on every section, because they describe the state of the whole thing rather than of one list. The first is your AI allowance, in the form Used about 2% of this month’s AI allowance, resets Sep 14. The second appears only when work is not being handed out: Agent runs are paused right now. Nothing is lost; they will resume automatically. When there is nothing to report, neither sentence renders at all.

The Inbox lists each proposed action with the agent that proposed it, the record it would affect, and exactly what would be sent or changed. Where the action carries more than the low risk tier, that word opens the sentence rather than sitting in a badge of its own. Nothing in the list has happened yet.

Two controls sit on the row itself, Approve and Reject, so clearing a queue does not mean opening eleven panels. The row itself is the third affordance: expand it for Details, or choose Edit before approving to change what would be sent before you agree to it. Editing an item creates a fresh preview, so what you approve is what you just read.

Each row carries its own clock, Expires today or 11 hours left, beside the agent’s name.

A decided row does not vanish. It re-renders in place with what happened to it, Queued, Running, Done or Rejected, and the header count goes down. An action that ran out of time before it was carried out reads Expired before it could run and offers Run it again.

Filter by agent, by record type or by risk to triage a busy queue; a filter that matches nothing says which filters are narrowing it. Arriving from a run shows the chip Filtered to one run with its own way to clear it. The full queue, its batch decisions and how a stale proposal is handled are covered in The Approvals Queue.

The Agents section opens with Results and the window it covers, This month. Four counts, and two of them open the inbox: Actions taken opens the queue, and Waiting on you opens it narrowed to what is still waiting. The two dispositions between them are counts you read, with nothing to press. The waiting count is the live queue figure the sidebar badge shows rather than a monthly one, so the tile and the badge can never disagree.

Under the four counts is a line per agent that did something this month. On a workspace where nothing has run yet the strip says so in words rather than painting four zeros: No agent activity yet this month.

Below that is the roster itself. Each row carries the agent’s name, how much it may do on its own in short form (Asks first, Routine work or Automatic), when it last ran, and a switch to pause or resume it. An agent that cannot run right now says why on its own row, in a sentence, with the one thing that fixes it beside it.

Activity lists runs across all your agents. Each row shows the agent name, what it ran on, its state, and when it ran. Filter by state to focus on what matters, or land here filtered to a single agent from its row, with a chip you can clear to see everything again. When nothing has run yet it reads No runs yet. Open a row for its timeline, or follow the chevron to the full trace.

A row offers something to press only where your click changes the outcome. While the platform is retrying on its own there is no button, because there is nothing for you to do. Where the cause is your allowance, your subscription, a permission or the agent’s own instructions, the button is the remedy rather than a retry.

Two of the eleven are about billing rather than about the agent. When the workspace has used its monthly AI allowance, every agent pauses itself and each of its runs offers the same button, Add Extra AI Credits. Extra AI Credits start the agents again before the cycle resets, and they resume on their own when it does. When the subscription itself needs attention, the row points at billing instead.

A test run is not a twelfth state. It wears the flag Test in front of whichever of the eleven it reached, so a dry run that found nothing reads Test · Done, nothing to do.

Opening a run gives you its full trace. A header summarises the run, with the agent, its state, and a flag when it was a test run or ran on an earlier version. A Stop run button appears while a run is still in flight. The trace then reads top to bottom as a story of the run.

Below the header, the run is laid out in sections:

• Pipeline replays the agent’s stages from the real run.
• What it read lists sanitized tool activity and public result summaries.
• What it decided shows a deterministic decision summary, never private reasoning.
• What it proposed lists each action, with its risk and what happened to it.
• Changes made shows the effects of any actions that were approved and ran.

Stopping a run is durable. It reaches work already handed out, prevents later claims, and reconciles any external effect that was already accepted. A stopped or finished run invalidates every proposal still pending on it, all at once.

An agent page has three tabs: Build, Runs and Settings. Which one you land on is decided for you. An agent that is live, or paused after having run, opens on its Runs, because what it did is what you came for. A draft, or an archived agent, opens on Build, because what it is is the only thing there is to see.

Naming a tab in the link always wins over that rule, and opening one run always lands you on Runs with the run open, so a link from a notification never hides the thing it pointed at. The tab is in the address, so it can be linked, reloaded and shared; flipping between tabs replaces rather than stacks, so Back leaves the agent instead of walking through the tabs you looked at, while closing a run you opened does take one Back press.

If your role cannot approve, agent proposals assigned to you still appear in My Work so you know what is waiting, and an approver on your team resolves them.