Lab notes from\nthe agent frontier.
A 6-module engineering lab on agent architecture — memory layers, planning loops, multi-agent coordination, eval design. Less marketing, more code. Every experiment runs in your OpenClaw sandbox so you can poke at the internals yourself.
Open the hood. Build the same agent loop six different ways. Measure what changes. The lab notes you wish someone had written for you.
Three real memory stacks: pure RAG, episodic, hybrid. Build all three, run the same workload, measure the actual differences. The numbers that make architects argue.
Five planner patterns from the literature, each implemented. The one that works in production vs. the one that wins the benchmark. Lab notes on the cost-of-replanning.
Four real coordination patterns. The one that works. The three that look good in a paper and fail in production. Build all four and find out why.
How to design an eval set that doesn't lie to you. Golden datasets, regression tests, the 'hard examples' problem. The eval philosophy.
The five unsolved problems in agent engineering as of 2026. What's been tried. What's worth trying. Pick one to explore in your sandbox.
Modules of hands-on lab work, end to end
Experiments that all run in your sandbox
Of guided lab work with measurements
Engineers shipped a working stack from this lab
Because real practice needs a real computer.
Most courses teach you to call an API, then leave you on your own. The OpenClaw Lab Course is built on OpenClaw because OpenClaw gives every student a real isolated machine: a browser, a terminal, a file system, and credentials to actual tools. You do real work, not pretend work.
A real linux box per lesson.
Every lesson in The OpenClaw Lab Course spins up your own private OpenClaw sandbox — a real Ubuntu machine with a real browser, real terminal, and a real home directory that persists. You write code that touches real files, opens real URLs, and ships real artifacts.
Your tutor IS an OpenClaw agent.
You're not watching videos. Lattice reads your code in real time, runs it in your sandbox, debugs alongside you, and asks the right next question. It remembers what you've already learned and what tripped you up last lesson.
Ship to your own stack.
By the time you hit Module 04 you're connecting your work to your real tools, your real data, your real environment. The course doesn't end with a toy demo — it ends with something you actually use, every day.
Six modules. One real thing shipped by the end.
Each module is self-contained but builds on the previous one. Start free with Module 01 — by Module 06 you're shipping from your own terminal.
The agent loop: dissected
Open the hood. Build the same agent loop six different ways. Measure what changes. The lab notes you wish someone had written for you.
- 01The minimal agent loop in 38 lines (no framework)
- 02Six variations: blocking, async, streaming, retry, parallel, batched
- 03What changes between them, measured: latency, cost, reliability
- 04When to pick which (a decision table you can trust)
- 05Lab: write your own loop from scratch and benchmark it
Memory architectures compared
Three real memory stacks: pure RAG, episodic, hybrid. Build all three, run the same workload, measure the actual differences. The numbers that make architects argue.
- 01Pure RAG: the simplest stack and where it actually fails
- 02Episodic memory: the 'remember the run' pattern
- 03Hybrid: cold storage + warm cache + working memory
- 04Eval: which one wins for which workload (and the answer isn't 'hybrid always')
- 05Lab: build all three and run the agent-eval suite on them
Planner experiments
Five planner patterns from the literature, each implemented. The one that works in production vs. the one that wins the benchmark. Lab notes on the cost-of-replanning.
- 01ReAct, Plan-and-Execute, Tree-of-Thought, Reflexion, AdaPlanner
- 02Implementing each in 60 lines
- 03The 'cost of replanning' equation that nobody publishes
- 04Why the simplest planner wins production 80% of the time
- 05Lab: implement one planner and run it on a real task suite
Multi-agent coordination patterns
Four real coordination patterns. The one that works. The three that look good in a paper and fail in production. Build all four and find out why.
- 01Centralised orchestration vs. peer messaging
- 02Shared memory vs. message passing
- 03Static role assignment vs. dynamic handoff
- 04Why most multi-agent demos collapse on the second iteration
- 05Lab: build a 3-agent team and break it intentionally
Eval design from first principles
How to design an eval set that doesn't lie to you. Golden datasets, regression tests, the 'hard examples' problem. The eval philosophy.
- 01Why most eval sets are lies (and what to do instead)
- 02Golden datasets: how to build one in an afternoon
- 03Regression tests for agents (the part nobody covers)
- 04Pareto-optimal eval design: cost vs. signal
- 05Lab: build an eval set for a real production agent
Open problems and what to try next
The five unsolved problems in agent engineering as of 2026. What's been tried. What's worth trying. Pick one to explore in your sandbox.
- 01Long-horizon planning: what's actually known
- 02Real-time tool use: where the gaps are
- 03Cost-optimal architectures: the open question
- 04Self-improving agents: hype vs. reality
- 05Lab: pick one open problem and run a 90-minute experiment
Watch a lesson run.
This is Module 02 · Lesson 04 — eval suite — Lattice walks a student through a real problem in their OpenClaw sandbox. Reads the code, runs it, sees the failure, shows the fix.
Four real things you'll have shipped.
Every module ends with a project that runs in your sandbox and produces a real artifact. Not toy demos — real working things you'll keep using after the course is over.
Module 01 graduation project
The first thing you'll ship in The OpenClaw Lab Course. A small but real artifact that proves the foundation lessons stuck. Runs in your OpenClaw sandbox by the end of week one.
A working artifact you'll keep using after the course is over — not a notebook, an actual tool.
Module 03 graduation project
The mid-course project. By Module 03 you've got enough of the core down to ship something with real moving parts. The Tutor reviews every commit.
A working system with real parts — the kind of thing you'd put in a side-project portfolio.
Module 05 graduation project
The advanced project. Combines what you learned in Modules 04 and 05 into one real-world build. The kind of thing senior engineers actually ship.
A senior-level artifact: the system you'd talk about in an interview as 'the thing I built that taught me X'.
Module 06 capstone
The final project. End-to-end, production-shaped, deployed from your sandbox. The Tutor stays online while you ship it.
Your portfolio piece. Deployed, observable, and small enough that you actually understand every line.
Meet your instructor.
Lattice is an OpenClaw agent that runs alongside you in your sandbox. It's not a chatbot bolted onto a video — it actually reads your work, runs it, sees the failures, and shows you the fix. Persistent memory means lesson 12 isn't lesson 1 all over again.
Reads your code in real time.
Lattice watches your sandbox file system and your terminal output. When you save a file it's already up to speed before you click 'help'.
Runs your code for you.
Lattice has its own shell in your sandbox. It can reproduce a bug in 4 seconds and tell you exactly which line is failing.
Remembers your past mistakes.
If you tripped over something in Module 02, Lattice will quietly catch the same pattern in Module 04 and call it out before you even hit run.
Adapts the pace to you.
If you're flying, lessons get shorter and the labs get harder. If you're stuck, Lattice slows down and adds practice exercises until the concept clicks.
Never says 'as an AI'.
Lattice has a personality — direct, dry, allergic to hedging. It tells you your code is wrong when your code is wrong, and tells you when you've done something clever without flattery.
Lives in your sandbox forever.
After the course ends Lattice stays — paste any code into your sandbox and it'll review the same way it reviewed your homework.
From engineers who shipped.
From engineers who came in skeptical and left with measurements they trusted.
“I'd been 'going to take a course' for 8 months. The first module is free and 90 minutes long. By minute 45 I had a working artifact in my sandbox. Lattice catching my dumb mistake in lesson 02 was the moment I realised this was different.”
“The OpenClaw sandbox is the unlock. Every other course had me copy-paste into a notebook that died when I closed the tab. Here I had a real linux machine with my own files, and Lattice was running my actual code. It feels like pair work with someone who's done this 1000 times.”
“Module 04 was the moment something clicked for real. I had a real artifact shipped by Friday afternoon. My team thought I'd been working on it for weeks. It was four lessons.”
“I almost quit in module 03. Lattice noticed I'd been re-reading the same lesson three times and offered me a different angle that fit how I think. Clicked instantly.”
“The thing nobody tells you about courses like this is how much production / real-world stuff most of them skip. Module 06 covers the parts that pay rent. That single module saved us from a surprise that would have cost a lot.”
“I run a 12-person team. We did the course as a group over 3 weeks. Every person shipped something real in their squad by the end. The ROI calc is going to look ridiculous when I write it up.”
Questions before you start.
Free to start. Pay when you ship.
Module 01 is free forever — no credit card. Pro unlocks the rest. Team gives every person their own sandbox + group progress.
Module 01 + your own OpenClaw sandbox + Lattice for the first 5 lessons. Most people finish this in an evening.
- Module 01 (1h+, 5 lessons + lab)
- Your own private OpenClaw sandbox
- Lattice for the duration of Module 01
- Discord access
All 6 modules of The OpenClaw Lab Course, unlimited sandbox time, Lattice for the full course, and Office Hours when you're stuck.
- All 6 modules
- Unlimited OpenClaw sandbox time
- Lattice for the full course
- Weekly Office Hours with the authors
- Lifetime updates as new modules ship
- Cert of completion + project showcase
For teams running The OpenClaw Lab Course together. Per-person sandbox, group progress, and a private Slack with Lattice.
- Per-engineer OpenClaw sandbox
- Group progress tracking dashboard
- Private Slack channel with Lattice
- Custom labs tailored to your stack
- Volume discount at 10+ seats
- Priority Office Hours
Stop reading about it. Ship something this week.
Module 01 of The OpenClaw Lab Course is free forever and most people finish it in an evening. Your sandbox spins up in 30 seconds. Lattice is online and waiting.
Start the course free