0.0. Course
What will this course teach you?
This course teaches the engineering discipline around an AI agent, not only the model call. You will inspect, run, and extend one completed Ops Copilot through the AgentOps lifecycle:
- Design an agentic loop and choose where deterministic code is better.
- Add typed tools, least-privilege Agent Skills, MCP, retrieval, workflows, and A2A.
- Enforce tests, behavioral evaluations, PII redaction, human approval, and append-only auditing.
- Put agentgateway in front of model, tool, and agent traffic.
- Package the application once and run it on local k3d or an optional GKE lab with kagent.
- Export telemetry through OpenTelemetry to self-hosted MLflow, Prometheus, and Grafana.
The result is production-shaped: it uses real boundaries and operational tools, while every action and dataset remains deliberately local and fictional. main is the finished reference, not an empty scaffold. The final capstone asks you to preserve its platform contracts while replacing the example domain with your own.
Who is this course for?
- Software engineers who can write Python and want more than a chatbot tutorial.
- AI and ML engineers who need repeatable evaluation, security, and delivery practices.
- Platform and SRE teams who will operate agent workloads and their data plane.
- Technical leaders who need to evaluate where agent autonomy creates value or risk.
You do not need prior ADK, Kubernetes, MCP, A2A, or GCP experience. You should be comfortable with Git, a terminal, and basic Python.
What reference will you use?
The Ops Copilot helps an on-call engineer investigate incidents for a fictional service. It can inspect an immutable SQLite seed, search logs and Markdown runbooks, load scoped instructions from Agent Skills, and propose or execute mock remediation.
State-changing tools require human approval. A write updates a disposable runtime database and appends an audit record in one transaction. The committed seed is never modified, so mise run data:reset always returns the lab to a known state.
Which learning path should you use?
| Path | Requires a model? | Requires cloud? | Purpose |
|---|---|---|---|
| Offline engineering | No | No | Install, inspect data, run static checks, and execute the deterministic test suite. |
| Required OSS path | Qwen3 through Ollama | No | Complete the agent, gateway, local platform, and observability outcomes account-free. |
| Optional provider | Gemini | Optional | Compare ADK's native provider adapter after the local path works. |
| Optional cloud lab | Gemini on Vertex AI | Yes | Practise GKE, Workload Identity Federation, GCS artifacts, and immutable delivery. |
The required path uses the Apache-2.0 open-weight Qwen3 model. It needs no account, no mandatory SaaS, and no usage fee; it does use your machine's compute, storage, and electricity. Chapter 2 connects directly to Ollama. Chapter 5 changes only the OpenAI-compatible base URL to route the same model through agentgateway.
How long should the course take?
Plan roughly 12 to 18 focused hours for the local path:
| Stage | Typical time |
|---|---|
| Overview and setup | 1-2 hours |
| Agent and capabilities | 4-6 hours |
| Quality and gateway | 3-4 hours |
| Platform and observability | 3-5 hours |
| Community and capstone planning | 1-2 hours |
Installation, image pulls, local model speed, and prior Kubernetes experience can change that substantially. A complete capstone implementation is intentionally open-ended and usually adds several focused sessions. The optional GKE lab adds provisioning time and should be completed separately.
How much does it cost?
The course content and code are free to use under their respective licenses. The offline and local Qwen3 paths have no model API fee, but they use your own compute and electricity.
Gemini and GCP usage can be billed. The optional GKE configuration targets a small, zonal, single-Spot-node lab and no public load balancer. The commonly stated under USD 20 per month target is conditional: it assumes your billing account still has the monthly GKE free-tier credit for one zonal cluster, light model/storage/network usage, and prompt teardown. Without that credit, the GKE management fee alone is USD 0.10 per cluster-hour. Check the current GKE pricing and the OpenTofu plan before creating anything.
A lab is not a production topology
One Spot node is interruptible and not highly available. The cloud chapter teaches identity, packaging, routing, persistence, and teardown. It does not claim to satisfy a production SLO.
Is the whole stack open source?
The required software stack is open source: Google ADK, agentgateway, kagent, MLflow, OpenTelemetry, Prometheus, Grafana, Ollama, the Apache-2.0 open-weight Qwen3 model, and this repository.
Gemini, Vertex AI, GKE, and GitHub hosting are proprietary services. They are optional integrations and are named as such. "Open-source course" means the course implementation is inspectable, modifiable, and self-hostable; it does not relabel a hosted model or cloud as open source.
How should you work through each page?
Every page is organized around practical questions. Run commands from the repository root unless a block says otherwise. Critical Python excerpts are included directly from source and checked during the site build. Complete each checkpoint before enabling the next prerequisite tier; for infrastructure chapters, verify the result and run the scoped teardown when you finish.
What is the first checkpoint?
Clone the repository and run the model-free test gate:
git clone https://github.com/MLOps-Courses/agentops-open-course.git
cd agentops-open-course
mise install
mise run install
mise run doctor
mise run test
You are ready for Chapter 1 when the base doctor passes and pytest reports that the enforced 95% branch-coverage threshold was reached. This stage does not require Ollama, a container engine, Kubernetes, GCP, or credentials.