Ch0 — The Dojo

“You cannot practice the sword in an empty field and call yourself ready for war. First, we build a war.” — Budo

The problem

Agent tutorials run against toy APIs, so the agents learn nothing and neither do you. We need a system that fails the way real systems fail: many services, real traffic, metrics, logs — and the ability to break it on purpose.

What you’ll build

The permanent lab for all ten chapters:

• kind cluster (3 nodes) — our “production”
• Online Boutique (Google’s microservices-demo) — 12 microservices, gRPC + HTTP, with a built-in load generator. Real traffic from minute one.
• kube-prometheus-stack — Prometheus, Alertmanager, Grafana
• Loki + Promtail — every pod log queryable
• Ollama — local model host, with a tool-calling-capable model
• A smoke test proving your local model can actually call tools

One command: just lab-up.

Prerequisites

Tool	Why	Check
Docker	kind nodes	docker info
kind ≥ 0.23	the cluster	kind version
kubectl	obviously	kubectl version --client
helm ≥ 3.14	observability stack	helm version
just	task runner	just --version
Python ≥ 3.11	agent code	python3 --version
Ollama	local models	ollama --version
Node ≥ 20	this site, later TS glue	node --version

Hardware honesty: qwen2.5:14b wants ~10 GB RAM/VRAM (Apple Silicon 16GB+ is comfortable). If tight, qwen2.5:7b works; llama3.1:8b is the fallback. See Appendix: Local Models.

Build

All files live in labs/ch00-dojo/.

1. The cluster

labs/ch00-dojo/kind-config.yaml defines 1 control plane + 2 workers with port mappings for Grafana/frontend access. Bring it up:

cd labs/ch00-dojo
just cluster

2. The shop

just shop

Deploys Online Boutique into namespace shop straight from the upstream release manifest, then waits for rollout. The loadgenerator deployment immediately starts simulating users — browsing, adding to cart, checking out. This traffic is the heartbeat every later chapter listens to.

Verify: kubectl -n shop get pods → 12/12 Running. Port-forward the frontend if you want to shop: just frontend → http://localhost:8080.

3. Eyes and ears

just observability

Installs kube-prometheus-stack and loki-stack (with Promtail) into namespace monitoring via Helm. Grafana at just grafana → http://localhost:3000 (admin / dojo-admin). Loki is pre-wired as a Grafana datasource.

Budo says: resist the urge to build dashboards. Your agents will query Prometheus and Loki directly — dashboards are for humans, and we are outsourcing the humans.

4. The mind

just models

Pulls qwen2.5:14b (primary) and llama3.1:8b (fallback). Ollama serves an OpenAI-compatible API at http://localhost:11434/v1 — this is the only reason our from-scratch chapters need zero provider-specific code.

5. Prove tool calling works

This is the step most people skip and then lose a day to. Local models vary wildly in tool-calling quality; verify yours before writing an agent around it:

just smoke

smoke_test.py asks the model “how many pods are running in the shop namespace?” with exactly one tool available (kubectl_get_pods), and asserts the model (a) chooses to call the tool, (b) produces valid JSON arguments, (c) uses the result in its answer. If smoke fails on qwen2.5:7b-class models, see the appendix before proceeding — do not start Ch1 with a model that fails smoke.

Break it

Even the dojo gets attacked. Kill the metrics pipeline and watch what your future agents would be blind to:

kubectl -n monitoring scale deploy/monitoring-kube-prometheus-operator --replicas=0

Now answer honestly: how long until you would notice? Restore it. Remember the feeling — Ch4 is about giving an agent better instincts than that.

Belt test

• just lab-up from zero completes without manual intervention
• kubectl -n shop get pods shows 12/12 Running
• Grafana shows traffic on the frontend service (loadgenerator working)
• {job="shop/frontend"} returns logs in Grafana → Explore → Loki
• just smoke passes 3 times in a row (tool-calling is consistently good, not occasionally)

What production would additionally need

Multi-AZ anything, persistent storage that survives kind delete, real ingress + TLS, RBAC beyond defaults. We skip all of it deliberately — the lab optimizes for “destroyable in one command.”