Build a multi-agent AI office, one guide at a time.
Thirteen guides, in dependency order, from a single agent-or-not decision through retrieval, guardrails, routing, and the flywheel that improves it. By the end you have a scoped, instrumented system, not a pile of demos.
Thirteen steps, in the order they depend on each other.
AI Agents vs. Prompts: When to Use Each, and When Not To
Use one prompt for single-step work you can check yourself. Use AI agents when the next step depends on what the model finds. Here is how I decide.
August 13, 2026
AI Structured Output: How to Stop Malformed JSON From Agents
A model's structured output looks fine until something downstream tries to read it. Here's what LinkedIn and Ramp built to catch it first.
September 3, 2026
Read Step 1 first.
- Scheduled
Tool Interfaces: Reaching a Database or an API
How an agent reaches out to fetch or change something: the MCP bridge, and the read-vs-write scopes that keep it from doing more than it was asked.
Shipping September 18, 2026
Read Step 1 and Step 2 first.
How RAG Answers From Your Own Documents
Learn how RAG searches your documents, where retrieval breaks, and how to choose chunks, embeddings, and a similarity floor.
August 8, 2026
How to Choose a Retrieval Cutoff
A practical way to set the similarity floor for RAG: real questions, source checks, ranked scores, and a refusal rule.
August 21, 2026
Read Step 4 first.
I Built a Knowledge Graph Next to My Vector Search
A knowledge graph and a vector search answer different questions. How I built a graph next to my vector search, and the query that shows which one you need.
September 14, 2026
Read Step 4 first.
AI Agent Routing: How Systems Send Work to the Right Specialists
AI agent routing makes one decision, cheaply: which specialist sees a job. Uber, RouteLLM, and Klarna show what getting that right, and wrong, costs.
September 4, 2026
Read Step 1 first.
What Parallel AI Agents Cannot See, and How to Fix It
Anthropic ran 16 coding agents on one bug and they kept overwriting each other. The mechanism, and the two fixes for a silent overwrite.
September 15, 2026
Read Step 7 first.
The AI Builder Stack, Translated: LangChain, LangGraph, CrewAI, n8n
Your proposal says LangChain, LangGraph, CrewAI, n8n, MCP, Streamlit. What each one is, when it is overkill, and what to ask before you sign.
August 18, 2026
Read Step 1 and Step 3 first.
AI Dialog State: How Chatbots Remember What You Said Mid-Conversation
Dialog state is how a chatbot remembers earlier turns. How Character.AI caches it cheaply, and why more history alone does not make replies reliable.
September 14, 2026
Read Step 1 and Step 9 first.
How to Evaluate an AI Agent Without Fine-Tuning Anything
How to grade an AI agent with a calibrated LLM judge, Cohen's kappa, and a pass@k reliability check, with no training run needed.
September 9, 2026
Read Step 4 and Step 5 first.
Multi-Tenant AI Systems: What Breaks When You Add Customers
What breaks when a multi-tenant AI product gets its second customer, in the specific places six companies already measured and fixed.
September 22, 2026
Read Step 8 and Step 11 first.
The Loop That Improves Itself: LoRA Fine-Tuning on Your Agent's Real Failures
The loop that improves itself: turn an agent's flagged failures into training data for a small, LoRA fine-tuned model, at NVIDIA's own numbers.
September 29, 2026
Read Step 11 and Step 12 first.
Tell me what you wish existed.
The form routes to what fits, or captures your need if nothing here does yet.