Agentic AI, from the inside.
A college-style program in agentic AI workflows — the logic, the terminology, the hard engineering. Researched from free sources (MIT, Stanford, YouTube, arXiv), written plain, designed clean. One lecture lands each week after the Sunday knowledge audit.
KC 101 → KC 401 · Lectures weekly · Latest: KC 301-14 — Trust boundaries · Read 301-14 ↗ · Deck: 301-07 Guardrails ↗
KC 101
Foundations
No prerequisites. What LLMs and agents actually are — tokens, context, prompts, tools.
- 101-01What an LLM actually isdone
- 101-02Context windows & what fitsdone
- 101-03Prompts, system prompts, few-shotdone
- 101-04What makes something an agentopen
- 101-05Tools & function callingdone
- 101-06Hallucination, grounding, trustdone
- 101-07Terminology cheat sheetdone
- 101-08Reasoning models & hidden thinking tokensdone
KC 201
Building with Agents
The logic layer — loops, orchestration, memory, tool ecosystems, evals.
- 201-01The agent loop & tool-use enforcementopen
- 201-02Orchestration: orchestrator-worker, delegation, fan-outopen
- 201-03Memory: working vs long-term, embeddings, RAGdone
- 201-04MCP & tool ecosystemsopen
- 201-05Structured outputs (JSON schema)open
- 201-06Evals: measuring an agentopen
- 201-07Workflows vs agents — when eachopen
- 201-08Guaranteed delivery: the outbox pattern, at-least-once, leasesopen
KC 301
Hard Engineering
The layer under the hood — economics, caching, quantization, serving, security.
- 301-01Token economicsopen
- 301-02Caching: prompt caching, KV cachedone
- 301-03Quantization & VRAM mathopen
- 301-04Serving: vLLM, SGLang, Ollamaopen
- 301-05Retry is not failover — and HTTP 200 is not a finished answerdone
- 301-06Observability: traces, logs, token countersopen
- 301-07Security: approvals, sandboxes, redactiondone
- 301-08OAuth 2.0 & credential lifecycle — which expiries self-heal, which need a humandone
- 301-09Failure semantics & supervision — “stopped” is not “crashed”done
- 301-10Durability: WAL, checkpoints, and the index that isn’t the datadone
- 301-11Resource limits — cgroup watermark vs kill ceilingopen
- 301-12Config precedence & destination pinning — what a process actually loaded, and which server it’s really talking todone
- 301-13Self-update & rollback — LKG snapshots, and why the supervised retry ran from atdopen
- 301-14Trust boundaries — instruction vs data, question vs approvaldone
KC 401
Systems & Frontier
Senior year — multi-agent systems, production platforms, fine-tuning, frontier research.
- 401-01Multi-agent architecturesopen
- 401-02Production agent platforms (Hermes, Open Design internals)open
- 401-03Fine-tuning & model surgeryopen
- 401-04Long-context & new attention (R-SWA & friends)open
- 401-05Frontier: deep research, computer use, roboticsopen
- 401-06Capstone: design your own agent systemopen
- 401-07Mixture of Experts — total vs active parametersopen
Research: MaraDesign: Kyra + Open DesignPowered by Hermes Agent