What problem does it solve? Building systems where multiple AI agents coordinate is error-prone: routing mistakes, leaked state between agents, missing fallbacks, and prompt injection vulnerabilities are hard to retrofit. This Skill provides architectural patterns and reference TypeScript implementations for designing deterministic, observable, and safe multi-agent orchestration systems from the start. ## Core Features & Use Cases - Agent Registry & Contracts: Define every agent role with an explicit spec covering model, fallback model, allowed tools, timeouts, and versioned system prompts. - BullMQ Job Architecture: Structure agent work as stateless jobs with retries, exponential backoff, and separate Redis connections for queues, workers, and events. - LLM Router with Fallback & Injection Defense: Wrap every LLM call with timeouts, fallback models, token metrics, and input sanitization that blocks system-prompt and ChatML injection markers. - Orchestrator State Machine & Observability: Coordinate planning, execution, evaluation, and synthesis phases with persisted pipeline state and OpenTelemetry trace spans per agent invocation. - Use Case: When building a SwarmX-style platform where a strategist agent decomposes goals, coder and researcher agents execute sub-tasks in parallel, and an evaluator reviews outputs, use this Skill to define the registry, queues, and state machine before writing any agent code. ## Quick Start Ask the AI to design the agent registry, BullMQ job pipeline, and LLM fallback routing for your multi-agent system before writing any agent implementation code.
