Agent Monitor — Live Ops
Nishad
🔶STUB — v-PROPOSED, fuller design to follow. This screen sketches the intent so no tile on Cortex Home is a dead door. Full session-drill-down and alert-routing behavior is a future REQ.

Agent Monitor — Live Ops

The live-ops view of the fleet — sessions, heartbeats, context-remaining, and stall/alert flags — distinct from Agent Manager's roster (who's assigned what). This is the productized version of the manual context-sentinel duty: catch a stall before it becomes a silent 26-hour hold.

Live sessions sessions · heartbeats · context remaining

AgentSession startedHeartbeatContext remainingStatus
opus-108:12 EST1s ago 61% ONLINE
opus-209:03 EST4s ago 29% CHECKPOINT DUE
codex-210:40 EST2s ago 74% ONLINE
ag-107:55 EST41m ago — STALLED

Stall / alert feed

🔴 ag-1 heartbeat stopped — 41m without a check-in, blocking DEV-1721 07-24 · 08:14
🟡 opus-2 context at 29% — checkpoint recommended before continuing 07-24 · 09:41
🟢 opus-1 resumed after compact — session healthy 07-24 · 07:02

Sentinel thresholds the rule this view enforces

35%Checkpoint trigger
~5 minHeartbeat poll
>15 minStale flag
>30 minStalled flag

What's on this screen

Built from: each agent's live session state, its last check-in, and how much working memory it has left. Stub screen — this is the intended data story; the fuller design (session drill-down, alert routing) is a future REQ.

On screenWhere it comes fromWhat it means to you
Live sessions tableeach agent's session start time and last check-inwhich agents are actively running right now
Context remaining barhow much of that agent's working memory is left before it needs to save progresswhen to nudge an agent to checkpoint before it forgets what it's doing
Status (Online / Checkpoint due / Stalled)whether an agent is active, running low on memory, or has gone silentwhether you need to step in
Stall / alert feeda running log of things that went wrong or recovered, newest firstwhat happened and when, without digging through logs
Sentinel thresholdsthe fixed rules this screen uses to flag an agentwhy an agent got flagged, in plain numbers