Build a project
Bring everything together: a real agent that does real work for you, with the production qualities from module 10.
The point of this track isn’t to read 11 modules; it’s to ship an agent that does useful work. This capstone is structured around picking a real problem you actually have and building the agent for it.
This module is being expanded with project templates and worked examples.
Coming in the next revision:
- Choose a project. Concrete examples by interest:
- Inbox triage agent — categorize, summarize, draft replies.
- Code review assistant — read PRs, comment on style/correctness, suggest improvements.
- Research assistant — multi-source search + synthesis on a topic.
- Documentation agent — keep internal docs in sync with the codebase.
- Operations agent — investigate alerts, propose runbook steps.
- Data analyst — natural-language Q&A over your warehouse.
- The architecture you’ll build: agent runtime → MCP servers (some pre-built, some custom) → vector memory → eval suite → observability.
- The non-negotiable production touches: cost ceiling, audit log, regression tests, human approval gate for any irreversible action.
- Worked end-to-end example. The inbox triage agent built from zero to running daily — code, deployment, eval.
This is where the track ends. If you completed the capstone you’re ahead of 90% of engineers building agents in 2026. The infrastructure is solved; the differentiation is the quality of the engineering. You’re now in a position to either ship an agent product or contribute meaningfully to one.
Where to go from here:
- The Agentic AI broad post on the main blog for a wider survey of the landscape.
- The AI/ML landscape map for adjacent tools beyond agents.
- The become a data scientist in 2026 post if you’re thinking about career direction in AI engineering.
- Send me what you built (or what blocked you) — the feedback shapes future modules.