Learn Multi Agent Engineering - Full Course (4+ Hours)

Build production-minded multi-agent systems with OpenAI Codex in this complete 4+ hour practical course. You’ll learn how to move from prompting to reliable agentic systems: prompt contracts, sequential and conditional workflows, parallel fan-out/fan-in, orchestration loops, checkpoints, subagents, evaluation, optimization, security, governance, and a complete capstone architecture. This course uses real Codex Desktop workflows and emphasizes evidence, reproducibility, bounded autonomy, human approval gates, safe deployment, and clear operational handoffs. What you’ll learn: • Design agent and workflow contracts with explicit goals, constraints, and done-when conditions • Build sequential, conditional, and parallel workflows without losing state or control • Orchestrate multiple specialist agents with bounded fan-out and reliable handoffs • Evaluate complete trajectories using deterministic checks and model-based judges • Diagnose failure modes, optimize the system, and add security and governance boundaries • Assemble a practical multi-agent release intelligence system in the capstone CHAPTERS 0:00 The Course Map: From Prompting to Agentic Systems Engineering 3:37 Codex Engineering Boundaries: Capability-First Agent Work 10:05 What Is an Agent? Model vs Agent vs Workflow 18:28 When Multi-Agent Is Worth It — and When It Is Overkill 24:16 The Agent Execution Loop in Codex Desktop 29:30 Prompt Contract: Goal, Context, Constraints, Done-When 33:51 A Tiny Controlled Loop: First Practical Codex Workflow 37:41 Codex Desktop as an Agentic Operating Environment 41:53 Threads, Context Windows, Compaction, and State 46:27 Plan Mode for Ambiguous Work 50:33 Goal Mode as a Long-Running Agent Loop 57:43 AGENTS.md as Durable Agent Memory 1:03:57 Sequential Workflows: A → B → C in Codex 1:09:01 Conditional Workflows: Branching on Evidence 1:17:51 Parallel Workflows: Fan-Out/Fan-In Without Chaos 1:22:11 Workflow Serialization: Make the Run Reproducible 1:26:12 Checkpointing and Resume 1:29:28 Workflow Debugging: Deadlocks, Missing Inputs, Bad Edges 1:32:18 The Orchestrator Loop: Select, Context, Run, Update, Stop 1:34:38 Termination Conditions: Budgets, Stop Text, Tests, Human Gates 1:36:27 Round-Robin as a Teaching Pattern 1:38:23 AI-Driven Orchestration: Let Codex Choose the Next Step Carefully 1:40:54 Plan-Based Orchestration and Intelligent Retry 1:43:23 Handoff Pattern: Moving Work Across Boundaries 1:45:26 Capability Discovery: Tell Users What This Workflow Can Reliably Do 1:47:07 Provenance: What Happened, Why, and With What Evidence 1:50:09 Interruptibility: Pause, Cancel, Resume, Recover 1:58:10 Dashboards for Agent State 2:08:11 Why Subagents: Parallelism, Context Isolation, and Role Specialization 2:13:29 Built-In Agents: default, worker, explorer 2:19:21 Custom Agents with Narrow Instructions 2:24:03 Bounded Fan-Out: One Agent Per Risk Lane 2:28:36 Steering, Stopping, and Closing Subagents 2:34:18 Evaluation-Driven Development for Agent Workflows 2:44:26 Trajectory Evaluation: Judge the Whole Run, Not Only the Final Answer 2:54:30 Deterministic Checks vs LLM Judges 3:04:31 The Cost of Multi-Agent Complexity 3:13:03 Ten Failure Modes of Multi-Agent Systems 3:25:03 Optimization Loop: Measure → Change → Re-Evaluate 3:35:22 When Agents Can Act: Why Security Becomes Design 3:43:31 Prompt Injection and Untrusted Context Boundaries 3:52:02 Audit Trails, Accountability, and Human Ownership 3:59:23 Responsible Deployment Checklist 4:06:43 Capstone Brief: Agentic Release Intelligence System 4:10:08 Build the Workflow Skeleton and State Model 4:12:10 Spawn Specialist Agents and Merge Evidence 4:14:04 Eval, Repair, Permissions, and Governance Pass 4:15:37 Final Dashboard, Handoff, and Architecture Review #OpenAI #Codex #MultiAgentSystems #AgenticAI