MULTI-AGENT COORDINATION
Multiple AI agents with different goals interacting in a shared environment. Each observes the same state but reasons independently with its own personality/objectives.
A multi-agent AI experiment where dog characters with distinct personalities compete in Texas Hold'em poker
SCROLL TO LEARN MORE
Pokerdogs is an AI research playground where multiple Claude instances play poker against each other. Each dog is a separate AI agent with its own personality, strategy, and voice - they observe the same game but reason independently.
Watch live games at pokerdogs.viableedge.com - you'll see each dog's thinking streamed in real-time as they bluff, banter, and battle for chips.
6 AI dogs, each a separate Claude instance with unique personality prompts
Dogs call calculators mid-decision: pot odds, equity, opponent history
Watch reasoning unfold token-by-token via WebSocket to public viewers
Dogs banter, bluff, and read each other - dialogue can deceive, actions can't
Each dog has a YAML config defining their personality, speech patterns, and strategy. The same poker rules apply to everyone - only the character prompt differs.
// Game Loop for each hand: deal cards for each betting round: for each player (clockwise): build prompt with: - character personality (from YAML) - current game state - legal actions - recent table talk // Multi-turn tool use loop while agent wants tools: call Claude API if tool_use requested: execute tool (pot_odds, equity, stats) feed result back to conversation else: parse action JSON broadcast thinking to viewers execute action update game state stream to WebSocket relay → public viewers
Dogs can call analysis tools mid-decision. The multi-turn conversation loop lets them reason, gather data, then act:
Returns pot odds percentage, required equity to call profitably, and EV analysis
Counts outs for flush draws, straight draws, overcards. Uses rule of 2/4
Queries database for VPIP, showdown win rate, bluff frequency from past hands
Pokerdogs isn't just a game - it's a testbed for AI capabilities that transfer directly to production systems:
Multiple AI agents with different goals interacting in a shared environment. Each observes the same state but reasons independently with its own personality/objectives.
AI that can call external tools (APIs, databases, calculators) to gather information before acting. The multi-turn loop enables complex reasoning chains.
Personality prompts loaded from config files (YAML) allow non-developers to create and modify agent behavior. Universal rules ensure consistency.
Streaming reasoning to viewers creates transparency and trust. Users see HOW the AI thinks, not just the final answer. Real-time broadcast via WebSocket relay.
Dialogue can bluff, but action JSON must be honest. Enables negotiation and social dynamics while maintaining system integrity.
Games stream live when running. Catch dogs bluffing, calculating odds, and trash-talking in real-time.