Agent activity

{ "workspace": "ministry-of-water", "status": "active" }

Show the work, keep people in command

Agents read, search, call tools and act for people. Every activity has a visible state, a plain-language status and a way to stop. Anything with consequences waits for a person.

The orb states for all of these are on Feedback and loading: searching for retrieval, solving for reasoning, connecting for tools, composing for writing, shaping for artifacts.

Reasoning

Chain of thought

Plans

• For work with three or more steps the agent shows its plan before starting, as a card with numbered steps. People can edit or reorder steps before they run.
• Steps that need approval are marked “Needs approval” up front, so nothing surprises the person later.
• While running, the current step shows the orb and progress; finished steps get a check; skipped steps are struck through with the reason.
• Stop is always available. Stopping keeps what is done and says what wasn't.

Queue

Tool calls

{ "workspace": "ministry-of-water", "status": "active" }

{ "workspace": "ministry-of-water", "status": "active" }

128 licences · 3 expiring before 30 Oct

{ "workspace": "ministry-of-water", "status": "active" }

403: the agent has no access to this workspace. Ask an administrator for the “Licences: read” role.

Confirmation

Tasks and background work

• Work longer than about ten seconds moves to a task: a row in the activity panel with a 32px orb, the title, the current step and elapsed time. People can leave and come back.
• When a task finishes: the orb is replaced by a check, a toast says what is ready (“Report ready · Open”), and the panel badge counts unseen results.
• Tasks touching files list them as chips; clicking opens the file at the changed place.

Checkpoints

Before an agent makes changes it can't easily undo, the conversation records a checkpoint: a dashed divider with the time and what follows. Restore returns files and the conversation to that point and says what will be undone.

Artifacts and previews

Developer surfaces

Coding and operations agents show terminals, test results, stack traces, file trees and environment variables. They follow the code style of the hub.

• Terminals use the dark code surface in both modes, the prompt in Oasis, success in green and failures in red with a symbol, never colour alone.
• Test results: one summary bar (passed, failed, skipped) and counts; failures listed first, each opening its stack trace.
• Stack traces collapse library frames and highlight the first frame in the customer's code.
• Environment variables and secrets are masked by default; reveal and copy are separate actions. See secrets.

Coverage

Dos and don'ts

Code

<details class="n-tool" open>
  <summary><n-orb state="connecting" size="20"></n-orb><code>admiral.licences.list</code>
    <span class="n-tag">Admiral MCP</span><span class="n-spacer"></span><span class="n-badge n-badge--info">Running</span></summary>
  <div class="n-tool__body"><p class="n-tool__label">Input</p><pre>{ "status": "active" }</pre></div>
</details>

<div class="n-confirm">
  <p class="n-confirm__title">…shield icon… Mai wants to send 1 email</p>
  <dl><dt>To</dt><dd>hosting@water.example.gov</dd></dl>
  <div class="n-confirm__actions">
    <button class="n-btn n-btn--secondary n-btn--sm">Don't send</button>
    <button class="n-btn n-btn--primary n-btn--sm">Send email</button>
  </div>
</div>

Last updated on 14 October 2026