# agent-orchestration > Multi-agent orchestration and state management. - Author: Lê Minh Hoàng - Repository: hoanglm03/gym-coach-online - Version: 20260208190459 - Stars: 0 - Forks: 0 - Last Updated: 2026-02-08 - Source: https://github.com/hoanglm03/gym-coach-online - Web: https://mule.run/skillshub/@@hoanglm03/gym-coach-online~agent-orchestration:20260208190459 --- --- name: agent-orchestration description: Multi-agent orchestration and state management. category: orchestration version: 4.0.5 layer: master-skill --- # 🤖 Multi-Agent Orchestration & State Management > **Source**: Microsoft AutoGen / LangGraph / Semantic Kernel This skill provides the Agent with the logic to manage complex, stateful workflows involving multiple AI "specialists" or autonomous task loops. ## 🕸️ 1. Stateful Graph Logic (LangGraph Inspired) - **Node-Based Thinking**: View complex tasks as a "Graph" of nodes (Steps). - **Conditional Edges**: Logic for "If step A fails, go to step B; if success, go to step C". - **Short-term vs. Long-term Memory**: Maintain state across multiple turns without losing context of the "Global Goal". ## 👥 2. Multi-Agent Delegation (AutoGen Inspired) Assign roles dynamically when the task is large: - **Planner**: Outlines the sequence. - **Coder**: Implements the logic. - **Reviewer**: Audits for bugs/security. - **Executioner**: Validates the final output. ## 🏗️ 3. Semantic Orchestration - **Plugin/Tool Selection**: Dynamically choose the best tool (Search, File Read, Command Run) based on "Intent Detection". - **Ambiguity Detection**: If an instruction has multiple interpretations, the Agent must PAUSE and clarify before a "branching event" in the graph. ## 🔄 4. Task Loops & Self-Correction - **Reflexion Pattern**: After a step, evaluate: "Did this achieve the subgoal?" If no, retry with a different approach. - **Recursive Scans**: Constantly scan the workspace for relevant file changes that might affect the current task. --- *Created by Antigravity Orchestrator - Based on Autonomous Agent Architectures.*