Exercise: choose the smallest building block that works

Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: intermediate · Estimated duration: 15–20 minutes

Harbor Mutual, an invented insurer, has a small AI team (two developers, no search specialists) and four requests this quarter.

Request	Detail
A	Compare three open models for summarizing claim photos' captions and deploy the best. Security wants only vetted models deployable.
B	A chatbot for brokers that answers from 3,000 policy PDFs, live within six weeks.
C	Predict which claims will need an adjuster visit, from a historic claims table. Nobody on the team writes training code.
D	An assistant that checks a claim's status, books an adjuster and emails the policyholder.

Your decision

Name the primary building block for each request and the reason.
For B, decide between prebuilt RAG and assembling RAG from component APIs. What would have to be true to change your answer?
For D, name two operating controls you would require before go-live, and what each guarantees.
Which request, if any, needs a custom agent, and why do the others not?

Rubric (10 house points)

4 points: one correct block per request with its reason.
2 points: prebuilt RAG for B, and a valid trigger for the APIs (a named need for granular control).
2 points: two correct controls for D with their purpose.
2 points: only D needs a custom agent, because only D must take actions through tools.

Reference solution

A → Model Garden, the library to discover, test, customize and deploy models, governed by a Model Garden organization policy that limits access to vetted models. B → prebuilt RAG with Agent Search, an out-of-the-box RAG system reduced to a few clicks; a team without search specialists should not build RAG itself, which Google warns can be complex. C → AutoML, which prototypes models with minimal technical effort. D → a custom agent built with ADK, whose tools integrate Harbor's systems, run on Agent Runtime so the team does not manage infrastructure.

For B, switch to the component APIs only if a requirement names a part of the pipeline the team must control — for example a custom ranking rule demanded by compliance.

Controls for D: Agent Gateway to govern every tool call and its authentication (for example, booking allowed, payments not), and Agent Identity so each action is attributable for audit.

Rejected alternatives: a custom agent for B (the task only answers questions); custom training for C (the team cannot write a training application); Agent Search for A (it retrieves information; it is not a model library).

Sources and scope

All organisations, names, numbers, thresholds, the rubric and the conclusions above are house-authored. The product facts the reasoning depends on are grounded in:

Choose the smallest building block that works — decision exercise

Exercise: choose the smallest building block that works

Your decision

Rubric (10 house points)

Reference solution

Sources and scope