Architectural Blueprints for AI Agents
summary
Nokia Networks builds agentic systems for private networks that target massive config complexity (example: a network function with 27,000 config entries) and a 2,000-person R&D org. Teams try LangChain LlamaIndex and Atomic Agent Frameworks then abandon generic stacks for a custom host/orchestration, toolbox and operation manual + ontology approach. Thousands of PDFs and Confluence pages fail RAG, so they implement property graphs (Neo 4G, kuzu) and a graph-RAG pipeline, demoed via a Jupyter notebook visualizer. Orchestration uses sequential, parallel+synthetizer and observer/evaluator patterns; memory strategies range from queues to LLM summaries, vector DBs and full graph memory.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:26AI agent challenges
- 3:16Agent integration demo
- 6:31Key learnings
- 8:44Defining AI agents
- 12:58Agent orchestration
- 16:33Architecture blueprint
- 21:50Operational manual
- 23:17Structured input/output
- 25:52Agent memory strategies
- 28:24Knowledge graphs for agents
- 36:00Tool integration
- 39:15Worker logic
- 40:40Q&A
related talks
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
25mBeyond Orchestration: An Open Source Architecture for Agentic Systems
31mAgentic AI architecture patterns
56mDesign Patterns for Multi Agent Systems
32mBuilding Multi-Agent Systems: From Concepts to Collaboration
40mHow To Tame Your Agents?
34mAI Agents: Transforming Autonomous Decision-Making
17mAI TechOps: Data and Infra for AI Foundation to Scale and Manage 1000s of Cohesive Agents