Live Agents

The Live Agents tab shows every data-classification agent that has registered with the console, with its connection status, software version, host details, and live health. Use it to confirm that the fleet is up, collect logs, drain an agent for maintenance, regenerate a key, or remove an agent.

Where: System Operations > Fleet Management > Live Agents

The list refreshes every 10 seconds, so you can leave the tab open as a live status board.

Overview

Agents perform the file extraction, linguistic coherence (LC) analysis, and PII and sensitive-data detection for the AI risk pipeline. Each agent sends a heartbeat to the console at a regular interval. The console records when it last heard from the agent, the version the agent runs, and the health metrics it reported.

Use this tab to confirm that the fleet is healthy before and during a classification run, and to act when an agent goes quiet or starts failing files.

Summary Tiles

Three tiles at the top of the tab give fleet-wide counts:

• Trino Agents — the number of registered Trino audit agents. Pipeline agents have no tile of their own.
• Online — agents that are sending heartbeats.
• Offline — agents with no heartbeat for 90 seconds. The console checks every minute.

The console's own entry (see Agent Cards) counts as Online.

If Online is lower than expected, find the offline agents in the list and read their offline reason.

Action Bar

• Collect All Logs — asks every online agent to push its logs to the console. A banner confirms how many agents were asked. The logs arrive within about 30 seconds, and the list then refreshes. Use it before you raise a support case.
• Refresh — reloads the list immediately.
• refresh in Ns · Nms — a countdown to the next automatic refresh, followed by the round-trip time of the last refresh in milliseconds, which helps you spot a slow console.
• N registered agents — the total number of registered agents.

Agent Cards

Each registered agent appears as a card. The top row of the card shows:

Item	Description
Display name	The agent's friendly name. You can edit it.
Type label	The agent kind, for example Pipeline · Extraction, Pipeline · LC, Pipeline · Classify, or Trino. The console's own entry is labeled Console. It is a placeholder for the console itself, not a worker, and it is always listed and always Online.
Status badge	Online (green), Degraded (amber), or Offline (grey).
DRAINING badge	Orange. Shown when the agent is in drain mode and accepts no new work.
Version	The software version the agent reports. A dashed no version pill shows when the agent has not reported a version yet. The images of pipeline agents appear in the All Agents grid and are upgraded on Agents Upgrade.

Below the top row, the card shows the agent's IP:port and hostname. It also shows the site label, with a pin icon, if one is set.

Health Pills (Pipeline Agents)

An online agent reports live health on each heartbeat. The card shows the values as a row of pills:

Pill	Meaning
CPU	The host's CPU percentage. Turns amber at 60% and red at 85%. The container's own figures are on the Agent Dashboard.
RAM	The host's memory percentage used. Uses the same color thresholds.
Disk	Free disk space inside the container.
Stage	The pipeline stage that the agent runs.
WIP	Files currently being processed.
60s	Files completed in the last 60 seconds.
N failed	Files that failed in the last 60 seconds. Shown in red, and only when greater than zero.
∑	Total files processed since the agent (container) started.
avg N ms	Average processing time per file over the last 60 seconds.

A persistent non-zero N failed value is the first sign of trouble. Collect the agent's logs.

Analyzer Pill

LC and Classify agents also show an Analyzer pill. It reports the state of the analyzer subprocess that runs inside the agent.

• ✓ up · N restarts · uptime (green) — the analyzer is responding. A climbing restart count means the analyzer is being restarted repeatedly. Investigate even when the pill shows "up".
• ✗ down · N restarts · last_death=... (red) — the analyzer is not responding and is being restarted with increasing delays. The pill shows the reason so that you can diagnose it.

Extraction agents have no analyzer pill, because their extraction component runs as a separate sidecar.

Health Pills (Trino Agents)

Trino agents show a different row of pills:

Pill	Meaning
Host CPU / JVM CPU	The host's CPU percentage and the JVM's own.
Host RAM / JVM Heap	The host's memory used and the JVM heap used.
Disk	Free and total disk space.
Disk I/O	Read and write latency in ms, IOPS, and utilization, or idle.
DB Connect ↑ / ↓	Whether the agent can reach ClickHouse.

Last Seen and Offline Reason

A clock line shows Last seen: <relative time> for online agents. For an offline agent, the line shows the offline reason in amber instead, usually the console's own Missed heartbeat — last seen …. Read it first when an agent drops off the fleet.

An agent that stays offline for 7 days is removed automatically and its key is revoked. On Kubernetes, an agent pod removes its own registration when it shuts down cleanly.

Log Push Status

Each agent shows the state of its log collection:

• Log push pending — waiting for agent heartbeat… — you requested logs and the agent has not yet delivered them.
• Logs last received: <time> — the console has received logs from the agent.
• No logs received yet — the console has never received logs from the agent.

Per-Agent Actions

The actions appear on the right side of each card.

• Edit (pencil) — edit the display name and site label inline. Click the green save icon to save or the ✕ to cancel.
• Regenerate key — creates a new API key for the agent and revokes the current one. The new key is shown once in a Copy this key now — it will not be shown again dialog. Reconfigure the agent with the new key.
• Remove (trash) — removes the agent's registration and revokes its key. A red confirmation bar asks "Remove ... and revoke its key? This cannot be undone." Click Remove to confirm or Cancel to go back.
• Request logs (online agents only) — asks this agent to push its logs now.
• Drain / Undrain (online agents only) — Drain stops new work from being assigned to the agent and lets running work finish. Undrain resumes normal assignment. Use Drain before you patch or reboot an extraction host.

Empty State

The console's own card is always listed, so the list is empty only if it fails to load. The tab then shows "No agents registered", with a hint to generate a key, install the agent, and wait for it to appear after its first boot. New agents register themselves on first boot and appear here after their first heartbeat.

Workflows

Check That the Fleet Is Healthy Before a Classification Run

1. Open Fleet Management > Live Agents.
2. Confirm that the Online tile matches your expected agent count, plus one for the console's own entry.
3. Check the health pills on each agent card for amber or red CPU or RAM, a non-zero failed pill, or a red Analyzer ✗ down pill.
4. For each Offline agent, read the offline reason on the clock line and bring the agent host back up.

Investigate an Agent That Is Failing Files

1. Find the agent whose card shows a red N failed pill or an Analyzer ✗ down pill.
2. Click Request logs on the agent card, or Collect All Logs in the action bar, to pull the logs to the console.
3. Read the offline reason or the analyzer last_death reason for the immediate cause.

Take an Agent Down for Maintenance

1. Click Drain on the agent card and wait for in-flight work to finish. The WIP pill drops to 0.
2. Patch or reboot the host.
3. When the agent is Online again, click Undrain to resume assignment.

Tips

• A no version pill on an otherwise online agent means the agent has not yet reported its version. It usually fills in within a refresh cycle or two.
• Regenerate key and Remove both revoke the existing key immediately. After Regenerate key, the agent stops authenticating until you reconfigure it with the new key. After Remove, the agent no longer authenticates.