Fetch ticket text and history
Domain 3 · 19% of the exam
Where the system
meets the world.
A model alone is a brain in a jar. Integration gives it hands, a library card, and a badge — and every one of those is a door that must be guarded, watched, and worth opening. The exam's heaviest domain.
Objectives
What you’ll be able to do
- Spot and fix capability bloat in tool/agent configs
- Find security gaps in authentication & authorization
- Justify accuracy–latency trade-offs
- Design monitoring and observability at scale
- Design RAG pipelines: chunking and indexing strategies
- Match retrieval strategy to data shape and query pattern
- Choose between MCP, API/CLI, and agent-to-agent
- Weigh progressive discovery vs monolithic context
The story
The robot with forty keys
Theo was proud of his new support agent. "It can do everything," he told Meera. "Read tickets, draft replies, issue refunds, change plans, delete accounts, query any database — forty tools." Meera looked at the tool rack the way a safety inspector looks at a forklift parked on a staircase. "And what do support staff actually do?" — "Read tickets. Draft replies." — "Then why does your robot carry the key to the demolition room?"
It wasn't only risk. With forty tools on its belt, the agent chose worse — every decision meant reading forty descriptions, and it kept reaching for exotic keys when the front door was open. They cut the rack to five. Accuracy went up. The security review shrank from a week to an afternoon. Capability bloat is a tax paid twice: once in risk, once in accuracy.
Then Meera introduced him to the librarian. "The binder problem, remember? You don't hand the model the library. You hire a librarian — retrieval — who fetches the three pages that matter. And how you cut the books into index cards —" she tapped the shelf, "— chunking — decides whether the librarian finds pages or confetti."
Metaphor map
Images that stick
| Concept | Image | Why it sticks |
|---|---|---|
| Tool use | The robot leaves the factory | The model steps outside to fetch and act. Every tool is a door — and a door swings both ways. |
| MCP | The universal port | One standard plug for every device. Build the adapter once; every client can use it. |
| RAG | The research librarian | Don’t hand over the library. The librarian fetches the three pages that answer tonight’s question. |
| Chunking | Cutting books into cards | Cut along chapters and you keep meaning; cut every 500 words and you make confetti. |
| Auth | Badges at the border | The agent wears its user’s badge. It opens exactly the doors its human could — never more. |
| Observability | The control tower | Every departure logged: who called what, with which arguments, costing how much, failing where. |
Tool rack
The tool rack
A support agent whose users only ever read tickets and draft replies. Toggle its tools and watch both meters — capability is never free.
Compose responses for human review
Move money to customer accounts
Modify customer subscriptions
Permanently remove user data
Query any table in production
Protocols
Three ways to plug in
Monolithic: load every tool definition and resource up front — simple, but the floor fills fast and choice quality degrades. Progressive: expose a thin catalog and let the agent discover details on demand — leaner context, better choices, at the cost of extra round-trips. At scale (dozens of tools, many servers), progressive discovery wins.
MCP
The Model Context Protocol: expose each system as a server once; any compliant client connects. Turns M×N integrations into M+N.
- Pick it for
- many tools, many consumers, cross-team reuse, ecosystem interop
- Watch for
- server sprawl — govern the catalog, version the contracts
API / CLI
Bespoke calls wired straight into your application. No protocol layer, no indirection — full control of every request.
- Pick it for
- one or two integrations, single consumer, tight latency budgets
- Watch for
- every new consumer re-implements the glue; drift accumulates
Agent-to-agent
Specialist agents exposing capabilities to each other — integration at the level of delegation, not function calls.
- Pick it for
- cross-boundary systems where each side owns its own agent and context
- Watch for
- compounding latency and error rates; trace across every hop
In production
A clinical-notes assistant that survived the audit
In production · healthcare
A clinical-notes assistant that survived the audit
- RAG
- Guidelines chunked by clinical section with specialty metadata; retrieval filters by department before similarity search.
- AUTH
- The agent inherits the clinician's credentials — it can never retrieve a chart its user couldn't open. No shared super-user key.
- MCP
- EHR, guidelines, and dictation exposed as MCP servers — one protocol, three systems, reused by every internal AI tool since.
- OBSERVE
- Every note traceable to its retrieved chunks and prompt version. When a guideline update broke retrieval, the trace found it in minutes.
Storyboard
The film, shot by shot
A brain in a jar can’t answer the door.
So we give it hands. Each tool, a door out of the factory. Each door, a risk.
Forty keys. Watch it hesitate. Watch the demolition-room key glint.
Five keys. The right five. Now it moves like it knows the building.
And for knowledge — the librarian. Books cut along their chapters, cards indexed, three pages fetched.
One port connects them all. And the tower sees every departure. Nothing moves unwatched.
Flashcards
Recall drill
Select a card to reveal the answer.
Quiz
Exam drill
Question 1
An internal reporting agent has read access to dashboards plus tools to modify pipelines, restart services, and drop caches. Analysts only ever ask it questions about metrics. Applying least privilege, what is the best change?
Question 2
A legal RAG system over 10,000 contracts answers clause-level questions poorly: retrieved chunks cut clauses mid-sentence and lose which contract and section they came from. Which redesign most directly fixes this?
Question 3
A platform team must connect Claude-based assistants to 15 internal systems, and other teams will reuse the same connections for their own AI tools. Which integration approach best fits?
Revision
Close your eyes and remember
Five keys, not forty. The agent wears its user's badge. Cut books along their seams. One port for many tools. And record every door that opens — you cannot debug what you did not trace.
Domain 3 · 19% of your exam — the single heaviest domain