How to Build Planning Agents without losing control
summary
Describes a working plan-and-execute pipeline built on an LLM compiler pattern with blueprint generator, planner, executor and joiner nodes implemented on LangGraph that turns a goal into subgoal decomposition, function calls to microservice-backed tools and a final report (example: an NVIDIA earnings-call workflow). Blueprints limit in-context tools so the planner stays small, MCP-style tool servers provide purpose and I/O contracts, and evals include component, end-to-end, LM-as-judge, code checks and human review. Avoids agentic workflows for fixed ETLs, safety-critical or low-latency cases.
AI-written summary · about
chapters (4)
related talks
40mHow To Tame Your Agents?
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
47mPlan Before You Build: Deterministic Planning Patterns for AI Agents
32mBuilding Multi-Agent Systems: From Concepts to Collaboration
34mAI Agents: Transforming Autonomous Decision-Making
43mArchitectural Blueprints for AI Agents
29mFrom Copy-Paste to Composition: Building Agents Like Real Software
02h 22mArchitecting and Testing Controllable Agents