Plan Before You Build: Deterministic Planning Patterns for AI Agents
summary
Five concrete planner patterns are demonstrated with runnable code and metrics: deterministic DAG workflows that run end-to-end in about 29 seconds and startups in ~400 ms (native image faster), HTN and GOAP for top-down decomposition and A replanning, utility AI for greedy exploration, and a multi-model council* deliberation that reduces hallucinations at the cost of higher API and latency; demos use Java 25, Spring Boot 4, Spring AI 2.0 M4, Vaadin Hilla 25, Cloud Run and Google Cloud Trace and they work as shown with observable traces.
AI-written summary · about
chapters (12)
- 0:00Introduction to agentic AI
- 1:11Agentic non-determinism
- 2:20Observability and LLM council
- 4:49The LLM council pattern
- 6:46Platform and design principles
- 12:12Planning patterns overview
- 15:04Workflow pattern exploration
- 20:21Supervisor LLM pattern
- 26:22Hierarchical task network
- 34:08Utility AI planning
- 39:33Goal-oriented action planning
- 44:08Consensus metrics and conclusion
related talks
08mReal AI Agents Need Planning, Not Just Prompting
35mAI Agent Control Logic with Pattern Matching
14mThe 4 Patterns of AI Native Development
47mAgentic AI Patterns
27mEmerging Patterns Agentic AI
34mAI Agents: Transforming Autonomous Decision-Making
54mGenAI Grows Up: Building Production-Ready Agents on the JVM
49mAgentic AI Patterns