Kill Switch

Kill Switch automatically detects and stops runaway AI agents before they burn your budget.

The Problem

AI agents can get stuck in infinite loops — making the same API calls repeatedly, running up costs without producing useful results. Common causes:

• Agent keeps retrying failed actions
• Circular reasoning in multi-step workflows
• Tool calls that always return the same result
• Prompt templates that generate identical requests

Traditional observability tools can only watch this happen. AgentGazer can stop it.

How It Works

Kill Switch uses SimHash — a locality-sensitive hashing algorithm — to detect repeated patterns.

Detection Signals

Signal	Weight	Description
Similar Prompts	×1.0	Prompts with Hamming distance < 3 after normalization
Similar Responses	×2.0	LLM responses with Hamming distance < 3
Repeated Tool Calls	×1.5	Identical tool/function call signatures

Scoring Formula

Score = similar_prompts × 1.0 + similar_responses × 2.0 + repeated_tool_calls × 1.5

When the score exceeds the configured threshold, the agent is automatically deactivated.

Prompt Normalization

Before hashing, prompts are normalized to detect semantic repetition:

Pattern	Replacement	Example
Numbers	<NUM>	"order #12345" → "order #<NUM>"
ISO Timestamps	<TS>	"2024-01-15T10:30:00Z" → "<TS>"
UUIDs	<ID>	"550e8400-e29b-..." → "<ID>"
Whitespace	Single space	Multiple spaces collapsed

This ensures that requests differing only in dynamic values are recognized as similar.

Configuration

Configure Kill Switch in the Dashboard: Agent Detail → Kill Switch Settings

Parameter	Default	Description
Enable	Off	Toggle to activate loop detection
Window Size	20	Number of recent requests to analyze
Threshold	10.0	Score threshold for deactivation

Tuning Tips

• Lower threshold = More aggressive detection (may have false positives)
• Higher threshold = More tolerant (may miss some loops)
• Smaller window = Detects short loops faster
• Larger window = Catches longer, slower loops

Recommended Settings

Scenario	Window	Threshold
Tight budget control	10	5.0
Balanced (default)	20	10.0
Tolerant	50	20.0

When Kill Switch Triggers

1. Agent is set to active = false
2. Agent's deactivated_by is set to 'kill_switch'
3. Event is recorded with event_type: "kill_switch"
4. Dashboard shows "Deactivated by Kill Switch" badge
5. Alert notification is sent (if configured)
6. Evidence payloads are archived for analysis

Incidents Page

The Incidents page provides a detailed analysis of all Kill Switch events.

Incidents List

Navigate to Incidents in the sidebar to see all Kill Switch events:

Column	Description
Time	When the Kill Switch was triggered
Agent	The affected agent
Provider	LLM provider in use
Score	Loop score vs threshold (e.g., "7.0/5")
Window	Detection window size used
Signals	Breakdown of detection signals (P=Prompts, R=Responses, T=ToolCalls)

Incident Detail

Click "View Details" to see the full analysis:

Scoring Breakdown

Visual breakdown of how the loop score was calculated:

Loop Score: 7.0 / 5.0 (140%)

[████████████████████████████░░░░░░|░░░░░░░░░░░░░░]
0                              Threshold          2x

The scoring table shows each signal's contribution:

Signal	Count	Weight	Score
Similar Prompts	3	×1.0	3.0
Similar Responses	2	×2.0	4.0
Repeated Tool Calls	0	×1.5	0.0
Total			7.0

Evidence Payloads

Collapsible list of the similar requests that triggered the Kill Switch. Each payload shows:

• Request body — The prompt sent to the LLM
• Response body — The LLM's response
• Character count — Size of each payload

This evidence helps you understand exactly what the agent was doing when it got stuck in a loop.

Reactivation

To reactivate a killed agent:

1. Go to Agent Detail page
2. Toggle the Active switch to On
3. The loop detector window is automatically cleared
4. Agent resumes normal operation

Manual Review Recommended

Before reactivating, investigate why the agent was looping. Check the Request Log for repeated patterns.

Alert Integration

Kill Switch events trigger alerts automatically. To receive notifications:

1. Go to Alerts page
2. Configure webhook URL and/or email for the agent
3. When Kill Switch triggers, you'll receive an alert with:
   • Agent ID
   • Loop detection score
   • Deactivation timestamp

API

Kill Switch configuration can also be managed via API:

# Get current config
GET /api/agents/:agentId/kill-switch

# Update config
PUT /api/agents/:agentId/kill-switch
{
  "enabled": true,
  "window_size": 20,
  "threshold": 10.0
}

Comparison with Other Tools

Feature	Langsmith	Langfuse	Helicone	AgentGazer
Loop Detection	❌	❌	❌	✅
Auto-Deactivation	❌	❌	❌	✅
SimHash Similarity	❌	❌	❌	✅

Other tools can only alert you about anomalies — you still have to manually stop the agent. AgentGazer stops it automatically.