Agent Guard

Built to solve the infinite loop problem in agentic workflows. Wraps your AI agents with circuit breakers, pattern detection, and stuck-agent analysis. You register each agent with thresholds, log actions as they happen, and call detect_loop or check_circuit_breaker before executing tools. It tracks the last 50 actions per agent, calculates diversity ratios, and flags exact repeats or low-progress patterns. When something goes wrong, get_stuck_report gives you token waste estimates and recovery suggestions. Useful when you're chaining LLM calls and need automatic intervention before an agent burns through your budget repeating the same failed action. Runs via stdio, integrates with Claude Desktop in one config block.