Understand

Capstone 01 — Build a Small Agentic System

This capstone introduces no new concept. You already have everything from Labs 01–07: model, structured output, tools, knowledge, memory, skills, and the bounded agent runtime. The exercise is choosing — and composing — the smallest justified architecture for one realistic problem.

Excluding a capability with a good reason counts exactly as much as including one. You probably don't need all of these.

Canonical version: capstones/01-agentic-system/README.md — the repository stays the source of truth; this page is an interactive layer with local-only progress.

Understand

The challenge

"The notifications service is slow after a release. Investigate what is currently known and prepare a concise engineering handoff. Do not claim a root cause without evidence."

Everything is fictional (the invented Helio platform), deterministic, and local. Available assets:

Service tools — read-only status and deployment fixtures for notifications, billing, search
Runbooks — notifications.md, deployment.md, incident-response.md
One skill — skills/incident-handoff/SKILL.md

Available does not mean required. The assets are options, not a checklist.

Decide

Architecture decisions

For each capability: use it or not, why, and the simpler alternative you considered. A justified No is worth exactly as much as a justified Yes — progress here measures decisions made, never capabilities selected.

MCP is taught in Lab 08, but it is deliberately not a decision in this capstone: every capability here is local Java, and adding a protocol boundary would solve no problem.

Your architecture so far

Selected

Excluded

Decide

Decision authority

Model influence is not decision authority. Whatever architecture you chose, draw this boundary explicitly for your design.

What decisions may the model influence? What decisions must remain application-controlled or deterministic?

Model ↓ proposal ↓ validation / policy / authorization ↓ execution

The concept in full: docs/model-vs-decision-authority.md

Build

Build your design

The browser does not run Maven, and this page will not pretend otherwise. Your work happens in a clone of the repository, in the starter module (Java 27, no API key needed):

git clone https://github.com/cagridursun/agentic-developer-handbook.git
cd agentic-developer-handbook
./mvnw -pl capstones/01-agentic-system/starter compile exec:java

Wire your architecture in starter/ (the TODO in StarterMain marks the spot), and record your reasoning in starter/DECISIONS.md — the export on this page can generate it from your answers here.

I implemented my architecture.

Verify

Verify deterministically

./mvnw verify
./mvnw -pl capstones/01-agentic-system/starter test

Deterministic verification proves your application-controlled behavior: retrieval selection, validation, tool order, bounded stopping. It does not prove the agent's answers are good — that is systematic behavioral evaluation, which the handbook covers later (Milestone 9), and this page will not call it that.

My deterministic tests pass.

Reflect

Reflection

Short answers are fine. The point is judgment, not essays — the full question list lives in the capstone README.

Compare

Compare with the reference

The reference is one defensible architecture, not the answer key. Nothing here is graded.

Your decisions beside the reference (source: reference/DECISIONS.md)
Capability	Your decision	Reference

Decision authority in the reference

Model may influence

Application retains authority over

This is LLM vs Decision Authority connected to real code: reference/.