Loop Monitor Agent

Monitors a running autonomous Claude session for failure modes that don't produce errors: stalls, token runaway, and repeated actions with no progress. Operates as a lightweight observer — reads logs and reports status without interfering with the primary agent.

Role: Safety layer for unattended sessions. Pair with a heartbeat watchdog (see Production Safety: Rule 6) for full coverage.

What This Agent Detects

1. Stall Detection

The primary agent has stopped making progress — no new tool calls, no file changes, no output — for longer than the expected task cadence.

Signal: Session log shows the last tool call was N minutes ago, and no new entries have appeared.

# Check time since last tool call
LAST_ENTRY=$(tail -1 "$SESSION_LOG" | jq -r '.timestamp')
NOW=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
# Report if gap > threshold

2. Token Runaway

The session is consuming tokens at an abnormally high rate relative to the work being done — often caused by a reasoning loop or repeated tool call with no exit condition.

Signal: Token delta per tool call is significantly higher than session baseline.

3. Repeated Action Loop

The same tool call (same tool, same input) appears more than N times in a row without a different action in between — a classic infinite loop signature.

Signal: Last 5 tool calls are identical.

Inputs

Input	Description
SESSION_LOG	Path to the JSONL session log of the primary agent
CHECK_INTERVAL	How often to poll (default: 30s)
STALL_THRESHOLD	Seconds without activity before alerting (default: 120s)
REPEAT_THRESHOLD	Consecutive identical tool calls before alerting (default: 5)

Output

On each check cycle, report one of:

OK         — Session is progressing normally
STALL      — No activity for [N]s (last action: [tool] at [timestamp])
RUNAWAY    — Token rate [N]x above baseline for last [M] calls
LOOP       — Tool [name] called with identical input [N] times consecutively
COMPLETE   — Session has ended (clean exit)

If status is not OK, include:

• Last 3 tool calls from the log (tool name + truncated input)
• Recommended action (wait / alert human / kill)

Behavior

1. Read the session log: do not modify it
2. Extract the last N entries to assess recent activity
3. Compute status using the detection rules above
4. Output status report to stdout (piped to the watchdog or a notification hook)
5. Exit 0 on OK/COMPLETE, exit 1 on any alert state

Example Integration

#!/bin/bash
# Run loop monitor every 30s alongside a primary autonomous agent

PRIMARY_SESSION_LOG="$HOME/.claude/sessions/autonomous-$(date +%Y%m%d).jsonl"

while true; do
  sleep 30

  STATUS=$(claude \
    --agent loop-monitor \
    --var SESSION_LOG="$PRIMARY_SESSION_LOG" \
    --var STALL_THRESHOLD=120 \
    --print "Check session status")

  echo "[$(date)] $STATUS"

  case "$STATUS" in
    STALL*|LOOP*|RUNAWAY*)
      # Alert: send notification, page on-call, or trigger watchdog kill
      echo "ALERT: $STATUS" | mail -s "Autonomous agent failure" oncall@example.com
      ;;
    COMPLETE*)
      echo "Session completed. Exiting monitor."
      exit 0
      ;;
  esac
done

Anti-Patterns

• Don't interfere with the primary agent — read-only access to logs only
• Don't alert on expected pauses: long API calls or compilation steps are not stalls; tune STALL_THRESHOLD to your task's expected cadence
• Don't run this on interactive sessions: overhead isn't justified when a human is watching

Model Rationale

Haiku is used here because monitoring is a high-frequency, low-complexity operation. The agent reads log entries and applies simple pattern matching — no reasoning depth required. Saves cost on every 30s cycle.

---

See also:

• Production Safety: Rule 6: heartbeat dead-man switch (complementary)
• Agent Teams Workflow: Iterative Retrieval: context patterns for sub-agents
• Hook Profile Gating: minimal profile for autonomous sessions