← All projects
AI AgentsOnline

Chess Coach AI

True multi-agent, not a scripted pipeline

The distinction I cared about here: the model has real decision authority. This isn't Python calling an LLM in a fixed order — it's a LangGraph ReAct loop where each agent reasons about what to do next.

The Critic is the piece I'm most attached to. Generated advice is frequently plausible and shallow, so a dedicated agent validates every claim and sends weak explanations back for revision before a human ever sees them.

What it does

  • Three specialized agents: Analyst (PGN patterns), Coach (personalized lessons), Critic (rejects weak advice)
  • Agents decide what to do next at each step — ReAct, not a fixed pipeline
  • Links a real chess.com account; pick how many games back and which time controls
  • Weights mistakes differently for fast time controls, where rushing is expected
  • Persistent session memory means you never re-authenticate
Next projectEvoke