Multi-Agent Orchestrator

Multi-Agent Orchestrator works through your pipeline before you write a single tool call. You describe your goal, available tools, and platform constraints; the skill interrogates the dependency structure of your task, identifies what is genuinely parallel versus what is a serial dependency in disguise, defines the minimum context each subagent needs, specifies output contracts and validation steps between stages, and maps where a failed or malformed agent response would cascade — and how to contain it. The result is an agent graph you can implement directly, not a list of vague suggestions.

A realistic input looks like this: 'I need to process 200 customer support tickets daily — classify intent, draft a reply, and log the outcome to a database. Running on Claude API, one developer, no existing orchestration framework.' The skill returns a structured architecture covering which of those three tasks can fan out in parallel, what each subagent's input contract must contain, how to validate the drafted reply before it reaches the logging step, and what the retry strategy is if the classification agent returns an unexpected label.

Sample output excerpt — Dependency Structure: Classification is a prerequisite for drafting; do not parallelise. Drafting and logging are independent once classification succeeds — fan out there. Per-Agent Brief (Draft Agent): receives {ticket_id, intent_label, raw_ticket_text}; returns {draft_reply: string, confidence: float}. Validation Gate: reject and retry if confidence < 0.7 or draft_reply exceeds 300 words. Failure Mode: if Draft Agent fails after two retries, route ticket to human_review queue and halt logging step.