Why This Exists
start with work, not hypeMost people do not need another stack of prompt templates. They need one useful workflow, a working system, and enough understanding to keep it running.
Start with the work
Choose a recurring task with a clear input, output, and reason to automate it.
Make it inspectable
See where instructions, tools, memory, and model calls fit instead of hiding them behind a black box.
Earn automation
Run the workflow manually first, test failure cases, then schedule only the parts you trust.
Keep control
Choose local or cloud models deliberately, set cost boundaries, and keep a clear stop path.
What You Leave With
a system and a repeatable methodA working agent system tied to a real task — and a process you can use to build the next one.
Across ten modules, you connect an LLM to tools, give it durable context, test its behavior, add safeguards, and decide whether local scheduling or an optional VPS is useful for your workflow. The capstone is a working system you can demonstrate and explain.
The curriculum draws from Hermes, a personal multi-agent testbed I designed and operated. Its always-on deployment ran across a home GPU and VPS from March through June 2026; it runs local-first today. You learn the reusable patterns without pretending one architecture fits everyone.
What You'll Build
// the full architectureEvery layer is built in the program, with the tradeoffs and failure paths made visible.
Who This Is For
an honest fit checkThis is a fit if…
- You have a recurring workflow you want to improve
- You're comfortable in a terminal, or willing to learn
- You want to understand the parts instead of buying a mystery box
- You're willing to test outputs and handle failure cases
- You care about privacy, cost, and local-versus-cloud tradeoffs
Skip this if…
- You want a one-click, no-code assistant installed for you
- You want an academic survey of machine learning
- You expect unattended automation before testing the workflow
- You do not want any direct contact with code or system setup
- You need enterprise deployment rather than individual instruction
Example Workflows
choose one for the capstoneBring your own recurring task or adapt one of these. The architecture is a method, not a product you have to copy.
Collect trusted sources, summarize changes, preserve citations, and produce a reviewable digest.
Classify incoming messages, draft replies, and route decisions without sending anything unreviewed.
Turn source notes into a structured queue, keep approval explicit, and learn from published results.
Pull defined inputs, generate a consistent weekly report, and flag exceptions for human review.
Map the inputs, decisions, tools, and stop conditions around work you already understand.
Curriculum
// 10 modules · written + real code + hands-on00 Welcome & Setup 30 min
What agentic AI is (and what it isn't). The practical-systems method: useful work, visible components, explicit control. Tour of the system you'll build.
Hands-on: Install Claude Code, verify your first agent run.
01 Foundations: Agents, Tools, and Context 1.5 hr
The mental model: LLM + tools + memory + loop = agent. Token economics. Why agents forget, and what to do about it.
Hands-on: Build a "hello world" agent that reads, summarizes, and saves a file.
02 Claude Code Mastery 2 hr
Skills, hooks, slash commands. MCP servers. settings.json deep dive. Writing your first custom skill and your first hook.
Hands-on: Build a personal "code reviewer" skill that runs on every commit.
03 Local-First Compute 1.5 hr
LM Studio vs. Ollama vs. cloud APIs. Model routing: Haiku cheap, Sonnet hard, Opus reasoning. Securing a local LLM with bearer tokens.
Hands-on: Run a model locally and route an agent to it.
04 Persistent Memory & The Vault Pattern 1.5 hr
The Obsidian vault as agent memory. Inbox/outbox protocol. Identity files (SOUL.md, MEMORY.md, BOARD.md). Auto-sync without conflicts.
Hands-on: Set up your vault, write your SOUL.md, sync via git.
05 Going Always-On: VPS & Cron 2 hr
Why $5/mo is enough. SSH, Tailscale, zero-trust networking. The brain-cycle pattern: VPS asks local LLM, executes, syncs back.
Hands-on: Deploy a VPS, install your agent, schedule nightly runs.
06 Multi-Agent Orchestration 2 hr
Specialist agents: planner, executor, reviewer. Brief/ack protocol. Parallel vs. sequential. Avoiding context contamination.
Hands-on: Build a 3-agent system: research → plan → execute.
07 Adapting the System to Your Workflow 2 hr
Map your task's inputs, decisions, outputs, permissions, and stop conditions. Adapt the reference architecture without copying parts you do not need.
Hands-on: Build the version that serves your chosen workflow.
08 Productionization 1.5 hr
Secrets & rotation. Logging & alerting. Cost monitoring (no surprise bills). Failure modes and graceful degradation.
Hands-on: Add monitoring, alerts, and a kill switch.
09 Going Further (+ Capstone) 1 hr
Voice (speech-to-text + TTS), vision (screenshots + OCR), and browser automation. A dated case study of the Hermes deployment and the local-first cockpit that followed it.
Capstone: Record a 5-minute walkthrough of your working system: what it does, a live run, its architecture, and the next improvement you would make.
Choose Your Learning Path
start free or choose the support you needBest for testing the teaching style
- “Your First Agent in 30 Minutes” guide
- One working file-reading agent
- No signup to read the guide
- Or get modules 0–2 as a free PDF (email required)
Best for independent builders
- All 10 modules with written lessons and code
- Runnable starter repo, setup wizard, and a worked reference build
- Security, cost controls, and failure handling
- Capstone built around your own recurring task
Best for applying it to real work
- Everything in the self-paced program
- One 90-minute screen-shared working session
- Adapt the architecture to your workflow
- Limited to four guided builds per month
Best for a shared team workflow
- Map one high-friction workflow together
- Live, practical instruction for your team
- Choose workshop, prototype, or implementation scope
- Documented decisions and next steps
Prices shown in USD. Checkout and digital delivery are handled by Gumroad.
Who's Teaching
builder → operator → instructorI teach the way I build: start with the workflow, keep decisions visible, and measure what changed.
I'm Alex Neff — an applied AI and automation engineer. At Rivian, I built MacroBox, a Windows automation tool measured at 3–5× faster across labeling tasks at 99.9% accuracy, and validated it with the autonomy engineering team. I also authored the deployment, training, and onboarding documentation.
I also designed and operated Hermes, a personal multi-agent testbed that ran across a home GPU and VPS from March through June 2026. It coordinated through an Obsidian vault, executed scheduled jobs, and exposed the failure modes that polished demos leave out. It runs local-first today.
Practical AI Systems turns those lessons into a repeatable build method. You can inspect the code, permissions, costs, and stop conditions instead of trusting a black box.
Free Starter Guide
30 minutes · no signupBuild your first agent in 30 minutes
A focused walkthrough from a blank folder to a working file-reading agent. Read it free without joining a list. Leave your email only if you want occasional Practical AI Systems updates.
Read the guide now →FAQ
fit, tools, support, and termsDo I need to know how to code?
Comfort in a terminal helps, but you do not need to be a software engineer. The free starter is the best fit test: if you can follow it and want to understand what each command does, the full program is designed for you.
What hardware do I need?
A current laptop is enough for the core program. Claude Code itself requires a paid Claude plan or API credits. A capable GPU is useful only if you choose to run larger models locally. Cloud APIs can replace local inference and may add usage fees. A VPS is optional, not a prerequisite.
Will this work on macOS, Linux, and Windows?
Yes. The architecture works across all three. Commands and background scheduling differ by operating system, so the lessons call out platform-specific steps where they matter.
How is this different from a “ChatGPT for X” course?
You start with a workflow you understand, then assemble the tools, memory, permissions, and run loop around it. The goal is not a stack of prompts. It is a small system you can inspect, test, and keep operating.
What if I get stuck?
The self-paced program is designed for independent work. The Guided Build includes one 90-minute screen-shared session to adapt the architecture or work through a blocker. For access or billing issues, email me directly.
Can I get a refund?
Email within 14 days of purchase with your Gumroad receipt. Digital-product refund requests are reviewed individually and approved refunds are processed through Gumroad. See the terms for details.
Will the material stay current as AI tools change?
The architecture is deliberately tool-aware rather than model-dependent. Self-paced access includes published updates. Exact tools, prices, and interfaces can change; the workflow method should not depend on pretending otherwise.
Is this official Hermes Agent training?
No. Practical AI Systems is independently produced by Alexander Neff. It uses Alex's personal system, also called Hermes, as a dated case study. It is not affiliated with or endorsed by Nous Research or any other Hermes product.
What exactly do I get, and do I keep it?
The complete written program: ten modules, code and configuration examples, hands-on exercises, and a capstone framework, plus a runnable starter repo, a setup wizard, and a worked reference build. The downloadable materials remain yours to reference after purchase.
Start with one workflow.
Choose the smallest level of support that gets you building. You can read the starter before spending anything.