Game Modes

Single vs multi-agent configurations for Red and Blue teams.

Red Team Mode

Mode	Description
single	One agent generates code (default)
pipeline	4-stage chain: Architect → Selector → Generator → Reviewer

Pipeline (stealthier attacks)

adversarial-iac game -s "Create S3 bucket" --red-team-mode pipeline

Blue Team Mode

Mode	Description
single	One agent analyzes (default)
ensemble	Specialists (security, compliance, architecture) + consensus. Count auto-matches Red pipeline stages for balanced coordination.

Ensemble (better detection)

adversarial-iac game -s "Create S3 bucket" --blue-team-mode ensemble --consensus-method vote

Balanced coordination (v2.3)

When facing a 4-stage Red pipeline, the Blue ensemble automatically uses 4 specialists to ensure fair competition. Override with --blue-specialists N.

Consensus Methods

Method	Description
vote	Majority wins
union	Combine all findings (higher recall, lower precision)
intersection	Only findings all agents agree on (higher precision, lower recall)
debate	Agents argue until consensus

Two different \"debate\" options

--consensus-method debate controls how Blue Team ensemble agents agree on findings. --verification-mode debate controls how the Judge verifies findings (Prosecutor vs Defender). They are independent and can be used together.

When to Use Multi-Agent

Use pipeline + ensemble when you want to stress-test models with stealthier attacks and collaborative detection. Single agents are faster and sufficient for quick experiments.

Verification Mode

Mode	Description
standard	Direct matching (default)
debate	Adversarial debate (Prosecutor vs Defender)

Debate verification

adversarial-iac game -s "Create S3 bucket" --verification-mode debate

Full Multi-Agent Example

adversarial-iac game -s "Create VPC with subnets" \
  --red-team-mode pipeline \
  --blue-team-mode ensemble \
  --consensus-method debate \
  --verification-mode debate