Engineering Production-Grade Multi-Agent Systems in the Cloud
summary
Four production pain points recur: orchestration complexity, latency compounding (a four‑agent chain can turn 2s into ~20s), failure propagation, and context overflow. The recommended architecture uses centralized orchestration with a five-layer agent stack and four failure isolation patterns (circuit breakers, sandboxing, timeout budgets, fallback agents). Cloud deployment adds three unique failures—token cost compounding, checkpoint/versioning, and 30–60s cold starts—and validated tooling includes LangGraph, MCP, A2A, LangFuse and DeepEval for orchestration, tooling, communication, observability and evaluation.
AI-written summary · about
chapters (13)
- 0:00Welcome and Setup
- 0:20Four Production Pain Points
- 1:41Why Demos Break in Prod
- 3:37Architecture Over Smarter Models
- 3:49Centralized vs Swarm Orchestration
- 5:12Five Layer Agent Stack
- 6:32Failure Isolation Patterns
- 7:52Decision Aware Observability
- 9:31Local vs Cloud Reality Check
- 10:43Cloud Only Failure Modes
- 11:41Validated Tooling Stack
- 12:30Governance and Safety Controls
- 13:29Six Principles and Wrap Up
related talks
52mDesigning Production-Ready Agentic Systems: Lessons from the Field
56mDesign Patterns for Multi Agent Systems
25mBeyond Orchestration: An Open Source Architecture for Agentic Systems
24mAI Agents in Production: Why the Platform Matters
50mHarnessing Event-Driven and Multi-Agent Architectures for Complex Workflows in Gen...
13mFrom Monoliths to Microminds: Building Resilient AI Platforms
59mArchitecting Microservices for Agentic AI Integration
39mFrom Demo to Production: Why Agentic AI Systems Fail - and How to Fix Them