The original title is "Two LangGraph Projects That Taught Me How to Design Multi-Agent Systems"

Short summary

A DevOps manager shares two LangGraph projects — a release-notes agent and a CI bottleneck optimizer — that taught him practical multi-agent design patterns. Key lessons include using router supervisors for clean extensibility, tuning interrupt models to match the type of human judgment each step requires, and enforcing structured JSON output on every LLM call for reliable routing. Both projects run on a lightweight stack with no GPUs or vector databases, proving agent systems can be simple and practical.

• •Router supervisor pattern lets you add sub-agents with one if/elif rule instead of a graph rewrite
• •Interrupt model must match judgment type: per-item for breaking changes, batch for standardized items, matrix view for trade-offs
• •Every LLM call should return structured JSON with a retry-plus-fallback strategy for parse failures

Generated with AI, which can make mistakes.