Beyond Orchestration: An Open Source Architecture for Agentic Systems
summary
Concrete reference architecture for open-source agent systems that claims to close the demo-to-production gap: durable agent controller with checkpointing and retries, Temporal and Celery for durable workflows, serving via VLLM and KServe (or RayServe), policy-as-code using Open Policy Agent, signed/versioned artifacts, token budget controls and tracing for PII and bias. Bold claims include 95% of work being safety/infra, 1–3 agent depth in practice, and rough estimate of ~five weeks per checklist item for two developers. Practical, tool-focused, not purely conceptual.
AI-written summary · about
chapters (11)
- 0:00Introduction to safe agents
- 1:38Reference architecture
- 4:17Agent controller design
- 6:48Serving and inference layer
- 9:22Tools as contracted services
- 10:56Bounded agent loops
- 12:40Action classification and HIL
- 14:29Policy and data governance
- 18:05Observability and efficiency
- 20:12Operationalization and closure
- 21:35Q&A session
related talks
56mDesign Patterns for Multi Agent Systems
10mKeynote: The Open Source Runway: Building the Foundation for the Agentic Era
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
33mWhen Vibe Coding Meets Agent Engineering: The Next Evolution of Software Architecture
43mArchitectural Blueprints for AI Agents
59mArchitecting Microservices for Agentic AI Integration
45mAn Agent Walks Into a Bar... and Meets Another Agent: Multi-Agent Systems in Action
44mBuilding Agentic AI Systems: From Transactions To Conversations