Agentic AI
Where most of my learning time has gone this year — past calling a chat completion endpoint, into what it takes to build agents that call tools, keep memory, and stay inside guardrails. Shipped as a real npm package, not just a repo of notebooks.
What I've Shipped
AgentCore
The problem: Every LLM provider ships its own SDK, message format, and streaming API, so switching models, or falling back to a second provider when one is rate-limited, usually means rewriting the integration layer.
Memorybook
The problem: Tools like NotebookLM are useful for asking questions over your own documents, but they are closed products, no self-hosted option, no visibility into how retrieval or memory decisions actually get made.
Skill Map
Roughly the order I actually learned this in — foundations first, then the parts of "agentic AI" that don't show up in a five-minute demo.
LLM Foundations
Tokenization, attention, embeddings, context windows — what actually happens between a prompt going in and a response coming back, not just which endpoint to hit.
Prompt Engineering
Zero/few-shot, chain-of-thought, ReAct, self-consistency, and getting output back as usable JSON instead of a paragraph you have to regex apart.
RAG & Vector Search
Chunking strategies, embedding models, Qdrant for similarity search, metadata filtering, and reranking — the pipeline most "AI features" quietly depend on.
Vectorless & Agent Memory
Where vector RAG breaks down — chunk-boundary loss, embedding drift, opaque similarity scores — and alternatives like PageIndex-style retrieval and LLM-generated wikis as agent memory.
Agentic Workflows
The perceive-decide-act loop, strict JSON tool schemas, parallel vs. sequential tool calls, and guardrails against an agent that won't stop calling itself.
Agent SDKs
The OpenAI Agents SDK and the Claude Agent SDK — multi-agent handoffs, managed tools, tracing — after building a CLI agent from scratch first to know what a framework is actually saving you.
AgentCore's Request Pipeline
A simplified trace of how AgentCore actually routes a request: guardrail in, gateway routing, model call, guardrail out.
- Pipeline idle. Waiting for a request.
Building something agentic?
Happy to talk through it — RAG that's actually accurate, an agent that needs real guardrails, or picking a model provider without locking yourself in.