Anomaly Detection

AgentWatch detects runaway agent loops heuristically, before the budget is even exhausted. This provides an early warning system that catches problems at iteration 4, not iteration 100.

How It Works

The AgentWatch edge engine maintains a rolling window of the last 5 iterations for every active session. On each ingest request, it calculates the token growth ratio between consecutive iterations.

The Quadratic Growth Signature

When an agent gets stuck in a context-appending loop, each iteration appends the previous output to the context window. This causes token counts to grow quadratically:

Iteration 1:  1,000 tokens
Iteration 2:  1,200 tokens  (1.2x growth)
Iteration 3:  1,440 tokens  (1.2x growth)
Iteration 4:  1,728 tokens  (1.2x growth)  ← ALERT FIRED

The detection algorithm identifies this pattern when:

1. At least 4 data points exist in the rolling window
2. The last 3 growth ratios are all > 1.2x and monotonically increasing

Why 1.2x?

The 1.2x threshold is calibrated to distinguish between:

• Normal variation — Token counts fluctuate ±20% between iterations (normal)
• Document processing — A one-time spike when processing a large document (not sustained)
• True runaway loops — Sustained quadratic growth from context accumulation

Plan Requirement

Anomaly detection requires a Pro or Enterprise plan. Free plans receive basic budget enforcement only.

Alert Channels

When an anomaly is detected, AgentWatch fires an alert via:

Slack Webhook

Configure SLACK_WEBHOOK_URL in your Worker environment:

wrangler secret put SLACK_WEBHOOK_URL

Alert format:

🚨 *Runaway Agent Detected* 🚨
Session `ses_8f2k9x3m` showed quadratic growth signature.
Recent prompt tokens: [100, 150, 220, 330, 495]

Deduplication

Alerts are deduplicated per session. Once an alert fires for a session, it won't fire again until the session expires from KV (24 hours). This prevents alert fatigue from a single long-running session.

Configuration

Anomaly detection is automatic for all sessions with 4+ iterations. No configuration required.

Custom Rules

For more granular control, you can create custom anomaly rules via the Rules API (Pro/Enterprise only):

curl -X POST https://api.agent-watch.dev/v1/rules \
  -H "Authorization: Bearer aw_live_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "block-high-growth",
    "condition": { "prompt_tokens_min": 50000 },
    "action": "block",
    "priority": 200
  }'

Limitations

• Detection requires at least 4 iterations per session
• One-time document processing spikes may trigger false positives (but won't block — only alert)
• The 1.4x threshold is tuned for context-appending loops; other failure modes may have different signatures