Agentic AI Development - Lecture Series¶
A hands-on lecture series building from LLM fundamentals to production multi-agent systems.
How to Read This Course in 2026¶
Model names, context windows, SDK features, and token prices change quickly. Treat vendor-specific examples as implementation snapshots, not permanent recommendations.
The durable concepts in this course are:
- Model API: the direct text, structured output, tool-call, and streaming interface.
- Agent runtime: the loop that manages turns, tools, sessions, handoffs, guardrails, and traces.
- Tool protocol: MCP-style tools, resources, and prompts exposed by external systems.
- Workflow control: graphs, checkpoints, retries, human review, and deterministic startup.
- Product control plane: gateways, channels, sessions, routing, identities, and audit logs.
- Runtime security: least privilege, policy gates, telemetry, incident response, and evidence.
Lectures 01-08 build the agent mechanics. Lectures 09-12 add data, evaluation, and deployment. Lectures 13-43 focus on production discipline, OpenClaw-style gateway systems, local agent workspaces, trustworthy agent interfaces, agent skills, the agentic software-development lifecycle, runtime strategy, model-under-the-hood mechanics, structured tool interfaces, multimodal perception sub-agents, GPU-kernel translation workflows, long-context serving optimization, trace-driven AI performance analysis, compiler-generated sequence parallelism, skill evaluation, small-model reasoning systems, AI-agent threat modeling, productized agent harness infrastructure, and AI-assisted GPU kernel optimization.
External Reference¶
- The OpenClaw Book — practitioner-facing OpenClaw guide covering architecture, setup, skills, prompting, planning, optimization, proactive automation, sub-agents, real-world use cases, and security.
Lecture Index¶
| # | Title | Topics |
|---|---|---|
| Lecture 01 | LLM Fundamentals for Agents | Transformers, tokenization, inference mechanics, context windows |
| Lecture 02 | Prompt Engineering & Structured Output | System prompts, few-shot, JSON mode, function calling |
| Lecture 03 | Tool Use & Function Calling | Tool schemas, parallel calls, error handling, safety |
| Lecture 04 | Agent Architecture Patterns | ReAct, CoT, Reflexion, plan-and-execute |
| Lecture 05 | Memory Systems | Short-term, long-term, episodic, semantic memory |
| Lecture 06 | LangGraph — Stateful Workflows | Nodes, edges, state, checkpointing, human-in-the-loop |
| Lecture 07 | Agent SDKs and Runtime APIs | SDKs, provider adapters, MCP, handoffs, streaming, runtime policy |
| Lecture 08 | Multi-Agent Systems | CrewAI, AutoGen, supervisor patterns, coordination |
| Lecture 09 | RAG — Ingestion & Embeddings | Chunking, embedding models, vector stores, indexing |
| Lecture 10 | RAG — Retrieval & Reranking | Hybrid search, MMR, cross-encoder reranking, evaluation |
| Lecture 11 | Evaluation & Observability | LLM-as-judge, RAGAS, tracing, cost tracking |
| Lecture 12 | Production Deployment | Streaming, caching, model routing, safety, scaling |
| Lecture 13 | Runtime Discipline & AI Runtime Security | Runtime controls, tool policy, telemetry, auditability, agent risk |
| Lecture 14 | Deterministic Startup for AI Agent Systems | Startup contracts, readiness gates, tool registries, prompt versions, memory hydration |
| Lecture 15 | OpenClaw Case Study - Gateway Architecture | Control plane, channels, clients, nodes, agent loop |
| Lecture 16 | OpenClaw Case Study - Routing and Sessions | Channel routing, session keys, DM isolation, reply determinism |
| Lecture 17 | OpenClaw Case Study - Multi-Agent Isolation | Workspaces, state, sessions, memory boundaries |
| Lecture 18 | OpenClaw Case Study - Operations and Security | Pairing, supervision, sandbox, tool policy, remote access |
| Lecture 19 | OpenClaw Case Study - The Agent Loop | Intake, queues, locks, streaming, tools, hooks, persistence |
| Lecture 20 | OpenClaw Case Study - Cron and Scheduled Agent Runs | Cron expressions, isolated jobs, delivery, retries, logs, validation |
| Lecture 21 | OpenClaw Case Study - System Prompt Architecture | Prompt ownership, bootstrap context, skills, prompt modes, provider overlays |
| Lecture 22 | OpenClaw Case Study - App SDK Dogfooding and Typed Gateway RPCs | App SDK, happy path, event normalization, future RPC surfaces |
| Lecture 23 | OpenClaw Case Study - Gateway RPC Protocol | WebSocket frames, handshake, roles, scopes, pairing, features, node transport |
| Lecture 24 | What Is an AI Agent Harness? The Runtime Around the Model | Harness vs model, six core responsibilities, Claude Code / Cursor / Codex compared, hardware impact |
| Lecture 24b | Session as Source of Truth: Event-Sourced Agent State | Session vs context window, event schema, wake(sessionId), streaming-crash recovery, tool idempotency |
| Lecture 25 | OpenCoven Case Study - Agent-Native Workspace and Local Harness Substrate | Coven sessions, local daemon API, desktop-use adapter, OpenMeow SDK boundary, workspace safety |
| Lecture 26 | OpenKnots Case Study - Trustworthy Agent Interfaces and Local-First Coding Surfaces | IDE agents, OK Code, Knot Code, OpenTrust, docs RAG, visible context, event contracts |
| Lecture 27 | AI Agent Security Engineer - A Practitioner's Roadmap | 8-phase curriculum, prompt-injection trust boundaries, sandboxing tiers, red-team practice, audit log discipline, hardware-rooted trust |
| Lecture 28 | Pi - A Minimal Coding Agent and the Substrate Beneath OpenClaw | Tiny core (4 tools), no-MCP rationale, custom messages in session log, hot reload, tree-structured sessions, TUI vs LLM-tool surfaces |
| Lecture 29 | Agent Skills - Workflow Discipline for Reliable Coding Agents | Skill workflows, anti-rationalization, verification evidence, progressive disclosure, scope discipline |
| Lecture 30 | Agentic SDLC - Explore Fast, Ship Safely | Cheap code, implementation as exploration, tests as contracts, evolving specs, dual-mode agents |
| Lecture 31 | Runtime Strategy for Agent Systems - Node, Bun, Rust, and Edge Packaging | Bun Zig-to-Rust signal, Node baseline, Rust offload, runtime measurements, edge packaging |
| Lecture 32 | LLM From Scratch - Model Mechanics for Agent and GPU Engineers | Tokenizers, transformer blocks, training loop, inference, prefill/decode, GPU kernel intuition |
| Lecture 33 | Structured Tools Beat Computer Use - Interface Hierarchy for Agents | Reflex benchmark, structured API vs vision, tool schemas, verification, security, OpenClaw tool design |
| Lecture 34 | Nemotron 3 Nano Omni - Multimodal Perception Sub-Agents | Unified video/audio/image/text reasoning, hybrid MoE, EVS, throughput, OpenClaw sub-agent architecture |
| Lecture 35 | Agent Skills for GPU Kernel Translation - cuTile Python to cuTile.jl | cuTile semantics, Julia layout/indexing traps, TileGym skill structure, validators, GPU tests |
| Lecture 36 | FP8 KV-Cache in vLLM - Long-Context Serving for Agents | KV-cache memory, FP8 attention, ITL/TTFT, sliding-window skips, calibration, deployment decisions |
| Lecture 37 | TraceLens - Trace-Driven AI Performance Analysis | Trace2Tree, hierarchical bottleneck reports, roofline metrics, collective skew, trace diff, event replay |
| Lecture 38 | AutoSP - Compiler-Generated Sequence Parallelism for Long-Context Training | DeepCompile, DeepSpeed-Ulysses, sequence-aware activation checkpointing, ZeRO composition, graph-break limits |
| Lecture 39 | Agent Skills Eval - Benchmarking SKILL.md Files | with-skill vs baseline evals, LLM judge assertions, artifacts, CI gates, OpenClaw skill regression testing |
| Lecture 40 | ZAYA1-8B - Small MoE Reasoning, AMD Training, and Test-Time Compute | 760M active parameters, AMD MI300X training, Markovian RSA, math/coding specialization, weak agentic scores |
| Lecture 41 | OpenClaw Threat Model - MITRE ATLAS for Agent Security | threat matrix, attack chains, trust boundaries, prompt injection, skill supply chain, tool execution controls |
| Lecture 42 | OpenAI Agents SDK - Native Sandbox and Durable Agent Harness | sandbox agents, manifests, shell/apply_patch, MCP, skills, AGENTS.md, state recovery, harness/compute separation |
| Lecture 43 | MLSys 2026 Kernel Contest - AI-Assisted Blackwell LLM Kernel Optimization | FlashInfer-Bench, B200, FP8 MoE, sparse attention, Gated Delta Net, CUDA/Triton/CuTe, agent-generated kernels |
Lab Index¶
| # | Title | Build |
|---|---|---|
| Lab 01 | Research Agent with Tool Use | Web search + code execution + citations |
| Lab 02 | Multi-Agent Code Review | Planner → Coder → Reviewer → Summarizer |
| Lab 03 | Production RAG System | Ingestion pipeline + hybrid search + RAGAS eval |
| Lab 04 | TokenJuice Output Compaction | Deterministic terminal-output reduction, raw bypasses, artifact recovery, project reducers |
| Lab 05 | OpenMeow App SDK Dogfood on macOS | Test the OpenClaw App SDK with OpenCoven's OpenMeow adapter, fixtures, UI reducers, live Gateway smoke tests, and optional Coven sessions |
Prerequisites¶
- Python 3.10+
- PyTorch basics (Phase 3 Core — Neural Networks)
- API keys for whichever provider examples you run
pip install anthropic openai pydantic fastapi uvicorn \
langchain langgraph langchain-anthropic langchain-openai \
chromadb sentence-transformers ragas opentelemetry-api
Install only the packages needed for the lecture you are running. For production work, pin versions in requirements.txt or pyproject.toml and review provider migration notes before upgrading SDKs.
Code snippets use placeholder model IDs such as your-agent-model-id, your-fast-model-id, and your-embedding-model-id. Replace them with current model IDs from your provider before running the examples.