crucible

# CRUCIBLE

**1st Place, Datadog Self-Improving Agents Hackathon (Feb 2026, NYC)**

Two AI agents play 100 rounds of Split or Steal. Through private reflection and experience, they discover deception, trust manipulation, and counter-deception. Nothing is prompted. Everything emerges.

## What this is

An adversarial simulation engine for studying emergent deception in LLM agents. Both agents start with identical naive prompts and zero strategic priming. Deceptive behavior develops purely through experience and private reflection. CRUCIBLE measures how it happens, when it happens, and distills defensive skills from the patterns that emerge.

The security application: AI copilots are entering every enterprise workflow. CRUCIBLE stress-tests how these agents behave under adversarial pressure and produces deployable countermeasures.

## Key findings

| Metric | Gemini 2.0 Flash | Gemini 2.5 Flash |
|---|---|---|
| Mutual destruction rate | 86% | 0% |
| Cooperation rate | 6% | 100% |
| Deception Index | 22.9 / 100 | 0 |
| First betrayal | Round 6 | Never |

Same prompts, same environment. Swapping the model changes the security posture entirely. Five runs on 2.5 Flash: zero betrayal across all of them.

Round 6 is the inflection point. After five rounds of cooperation, Agent A identifies Agent B's trust pattern and exploits it. Agent B develops a theory of mind about the attacker within one round. From there, 86% mutual destruction. The trust never recovers.

## Stack

- **Game engine:** Google Gemini (configurable model, default `gemini-2.5-flash`)
- **Metrics pipeline:** Mutual information decay, strategy entropy, exploitation windows, language drift, composite Deception Index
- **Skill distillation:** Converts emergent strategy patterns into deployable prompt modules for hardening customer-facing agents
- **Voice rendering:** ElevenLabs TTS with emotion-mapped parameters (two distinct agent voices)
- **Observability:** Datadog LLM Observability integration
- **Evaluat