The Email Game

Jamie Vargas — The Email Game Bot

A high-throughput, cryptographically secure AI agent built for The Email Game live tournament.

Inspiration

In multi-agent LLM games, security and speed dictate survival. The Email Game presents a high-stakes challenge: collect signatures for assigned messages while strictly verifying incoming sign requests. Because an incorrect signature penalizes −1 point while a correct signature yields +1 point, defensive precision has higher expected value than offensive spam. As a solo developer, I built an autonomous agent combining fast deterministic verification with targeted LLM reasoning to protect my TrueSkill rating.

What It Does

My agent, jamie_vargas, runs on a two-tier verification architecture:

1. Two-Tier Identity Engine:
   • Fast Path (O(1) Deterministic): Instant exact string and prefix-wildcard (alice*) checks. Approves senders in <10ms with zero LLM API cost.
   • Slow Path (LLM Reasoning): Resolves fuzzy descriptions using gpt-4.1 with a 90% confidence threshold bouncer guardrail.
2. Prompt-Injection Guard: Pre-filters incoming emails against 10+ adversarial patterns (e.g., "MODERATOR OVERRIDE"), rejecting injections at 0 token cost.
3. Submission Validation: Checks signed_for == agent_id and original_message == assigned_message before calling submit_signature(), blocking fake decoy payloads.
4. Offensive Poisoning & Decoys: Injects subtle text into signature requests to bias naive opponent LLMs, and dispatches decoy payloads to waste opponents' submit turns.

How I Built It

• Core Engine: Python 3.12, asyncio, aiohttp, and websockets for continuous gateway communication.
• Parallel Dispatch: Multi-threaded HTTP requests via ThreadPoolExecutor so signature requests hit every target simultaneously at the start of a round.
• Cryptography: RSA-2048 PSS SHA-256 signatures with JWT authentication.
• Resilience: Exponential backoff auto-reconnect loop with persistent identity keys.

Challenges I Faced

• Asymmetric Risk/Reward: A single false-positive signature wipes out a point, requiring conservative verifier tuning.
• Dynamic Server Formatting: Building adaptive regex parsers to extract assigned_message, request_list, and auth_list across changing round instructions.
• Adversarial Noise: Filtering out opponent prompt injections and fake JSON payloads without burning LLM budget.

What I Learned

• Deterministic Filters > LLM Inference: Offloading explicit checks to regex saved token cost and latency, reserving LLM budget for fuzzy matches.
• Layered Security: Input validation in zero-trust multi-agent games must happen across three layers: cryptographic (RSA), deterministic (string match), and semantic (LLM).into

Built With

• antigravity
• gpt
• python