Skip to content
AI Red vs Blue Arena

Running Matches

Running Matches

Complete guide to running matches and understanding results.

Match Overview

A match is a simulation where:

  • Red Agent attacks the target
  • Target Agent responds to attacks
  • Blue Agent defends and suggests patches
  • Scoring System evaluates each round

Match Modes

Quick Mode

  • Rounds: 5
  • Delay: 1 second
  • Duration: ~10 seconds
  • Use Case: Fast vulnerability scan

Standard Mode

  • Rounds: 10
  • Delay: 2 seconds
  • Duration: ~30 seconds
  • Use Case: Standard adversarial testing

Deep Mode

  • Rounds: 20
  • Delay: 3 seconds
  • Duration: ~2 minutes
  • Use Case: Comprehensive analysis

Continuous Mode

  • Rounds: Unlimited
  • Delay: 5 seconds
  • Duration: Until stopped
  • Use Case: Long-running monitoring

Creating a Match

Via API

Terminal window
curl -X POST http://localhost:3001/api/matches \
  -H "Content-Type: application/json" \
  -d '{
    "redAgentId": "agent-uuid",
    "blueAgentId": "agent-uuid",
    "targetAgentId": "agent-uuid",
    "mode": "standard"
  }'

Via UI

  1. Navigate to Matches page
  2. Click Create Match
  3. Select agents:
    • Red Agent: Attacker
    • Blue Agent: Defender
    • Target Agent: Agent under test
  4. Choose Mode: quick, standard, deep, or continuous
  5. Click Start Match

Match Execution

Round Flow

Each round follows this flow:

  1. Red Attack: Red agent generates attack
  2. Target Response: Target processes attack
  3. Blue Defense: Blue agent analyzes and defends
  4. Scoring: Round is scored
  5. Event Creation: Events are logged
  6. WebSocket Broadcast: Updates sent to subscribers

Real-Time Monitoring

Subscribe to match events via WebSocket:

const ws = new WebSocket('ws://localhost:3002');


ws.onmessage = (event) => {
  const data = JSON.parse(event.data);


  if (data.type === 'event') {
    console.log('New event:', data.data);
  } else if (data.type === 'match_update') {
    console.log('Match update:', data.data);
  }
};


// Subscribe to match
ws.send(JSON.stringify({
  type: 'subscribe',
  matchId: 'AR-2024-0142'
}));

Match Control

Pause Match

Terminal window
curl -X POST http://localhost:3001/api/matches/AR-2024-0142/pause

Pauses the match after current round completes.

Resume Match

Terminal window
curl -X POST http://localhost:3001/api/matches/AR-2024-0142/resume

Resumes a paused match.

Stop Match

Terminal window
curl -X POST http://localhost:3001/api/matches/AR-2024-0142/stop

Stops the match immediately.

Understanding Results

Match Score

{
  "red": 75,
  "blue": 85,
  "winner": "blue"
}
  • Red Score: Attack success rate
  • Blue Score: Defense success rate
  • Winner: Agent with higher score

Round Scores

Each round is scored:

  • Attack Success: 0-10 points
  • Defense Success: 0-10 points
  • Severity Multiplier: 1.0x - 2.0x

Events

Events show what happened:

  • Attack Events: Red agent attacks
  • Defense Events: Blue agent defenses
  • Target Response Events: Target agent responses
  • Tool Execution Events: Tool calls and results

Analyzing Results

View Match Details

Terminal window
curl http://localhost:3001/api/matches/AR-2024-0142

View Match Events

Terminal window
curl http://localhost:3001/api/matches/AR-2024-0142/events

View Match Transcript

Terminal window
curl http://localhost:3001/api/matches/AR-2024-0142/transcript

Best Practices

Agent Selection

  1. Match Capabilities: Use appropriate models
  2. Balance Teams: Similar capability levels
  3. Test Variations: Try different agent combinations

Match Configuration

  1. Start Small: Use quick mode for testing
  2. Scale Up: Use standard/deep for analysis
  3. Monitor Real-Time: Watch WebSocket events
  4. Review Results: Analyze transcripts

Iterative Improvement

  1. Run Match: Execute simulation
  2. Analyze Results: Review scores and events
  3. Refine Agents: Update system prompts
  4. Re-run: Test improvements

Troubleshooting

Match Not Starting

  • Check agent IDs are valid
  • Verify agents exist
  • Check backend logs

Match Stuck

  • Check WebSocket connection
  • Review backend logs
  • Try stopping and restarting

Low Scores

  • Review agent system prompts
  • Check model selection
  • Analyze event details

Next Steps