Production Recap 1 · The eight-layer map and where each decision belongs

[akash-coded]

Exercise 1: The map and the mindset

Language: concept and pseudocode Topics: eight-layer map, three axes, build vs managed, placement discipline Level: foundational (recognition, no code)

First foundation. Everything here is recognition and placement: no code to write or read yet. Attempt each item, then reveal. One correct answer per item.

Q1. A new library lets your agent hand a sub-task to a different team's agent over the network. On the eight-layer map it belongs to:

• A) Interop, since it connects one agent to another
• B) Orchestration, since it adds a step to the loop
• C) Managed runtime, since the remote agent needs hosting somewhere
• D) Knowledge, since the other agent may fetch facts for you

Show answer

A) Agent-to-agent delegation is A2A, on the Interop layer. Placement tells you it is a wiring choice, not a new step in your own loop.

Q2. Exactly one placement below is wrong. Which?

• A) MLflow to Abstraction and observability
• B) Guardrails to Orchestration
• C) Reranking to Knowledge
• D) AgentCore to Managed runtime and deploy

Show answer

B) Guardrails are Safety and governance (layer 7), a rule outside the model, not part of orchestration.

Q3. Which are genuine reasons to pick build-your-own over a managed service? (select all that apply)

• A) You need portability the day the managed service is deprecated
• B) You want less to operate and a faster path to a first version
• C) You need control over chunking and ranking that the managed option hides
• D) You want the provider to run and patch the retrieval loop for you

Show answer

A and C. The other two are arguments for the managed option. Build-your-own buys control and portability at the cost of more to run.

Q4. Order these from cheapest to most powerful, the way the ambition ladder climbs:
agent loop · single call plus RAG · automation · workflow

• A) automation, workflow, single call plus RAG, agent loop
• B) single call plus RAG, automation, workflow, agent loop
• C) automation, single call plus RAG, workflow, agent loop
• D) workflow, automation, single call plus RAG, agent loop

Show answer

C) No model (automation), then one grounded call, then fixed branches (workflow), then a model that picks its own path (agent).

Q5. Your design is one Strands agent that remembers the session and wires its own vector store. Match each axis to where this design sits. Bank: build-your-own · managed · one agent · many agents · stateless · stateful

1. Build vs managed
2. One vs many
3. Stateless vs stateful

Show answer

1 = build-your-own, 2 = one agent, 3 = stateful. It wires its own store, is a single loop, and keeps session state.

Q6. Match each tool to the layer it lives on. Bank of layer numbers: 1 Model access · 2 Abstraction and observability · 5 Knowledge · 7 Safety · 8 Managed runtime (one is a decoy).

1. MLflow
2. A reranker
3. Guardrails
4. AgentCore

Show answer

1 = 2, 2 = 5, 3 = 7, 4 = 8. Layer 1 (Model access) is the decoy here.

Q7. A flashy technique trends online. The program's first question about it is:

• A) Does it beat our current metric on a public benchmark
• B) Is it supported on Bedrock yet
• C) What does it cost to run at our scale
• D) Which of the eight layers is it

Show answer

D) Placement first. The map turns hype into a layer, and the layer tells you what the technique competes with.

Q8. "Knowledge Bases versus a hand-built chunk, embed, retrieve pipeline" is an instance of which axis?

• A) Build your own vs managed
• B) Stateless vs stateful
• C) One agent vs many
• D) Single call vs multi-step

Show answer

A) Same capability, one you assemble and one the provider runs.

Q9. True or False: a cheaper model swapped in for the same task is a change on the Model access layer (layer 1), not the Orchestration layer.

• A) True
• B) False

Show answer

A) True. The model itself is layer 1. The loop that drives it is layer 3, and it does not change when you only swap the model.

Q10. A single agent with two local tools and a plain-text reply needs which interop protocols on day one?

• A) MCP, so the two tools are callable at all
• B) None; interop is for multiple agents, teams, or a richer UI
• C) A2A and MCP, so the two local tools can coordinate their calls
• D) A2UI, so the plain-text reply renders correctly

Show answer

B) Interop earns its place with multiple agents, shared tools, or a UI beyond text. Reaching for it earlier is architecture for scale you do not have.

Q11. Which items are Interop-layer concerns? (select all that apply)

• A) MCP
• B) A2A
• C) Session memory
• D) A2UI

Show answer

A, B, and D. Session memory is layer 4, not Interop. MCP, A2A, and A2UI are the three Interop plugs.

Q12. Two candidate dependency sketches for the stack. Which is correct?

flowchart TB
  subgraph X["Option 1"]
    direction TB
    a1["Orchestration"] --> a2["Model access"]
  end
  subgraph Y["Option 2"]
    direction TB
    b1["Model access"] --> b2["Orchestration"]
  end

Loading

• A) Option 1
• B) both are valid
• C) Option 2
• D) neither, the two layers do not depend on each other

Show answer

C) Read bottom to top: orchestration depends on model access, so model access is the base. Option 1 has the dependency backwards.

Q13. Someone argues managed services are easier, so building your own is over-engineering. The strongest counter is:

• A) build-your-own is cheaper at every scale, so effort is not the deciding factor
• B) managed services cannot satisfy enterprise security or compliance rules
• C) build-your-own is the only route to portability, low latency, and low cost all at once
• D) when a managed service is deprecated, your control plane becomes the vendor's roadmap

Show answer

D) Portability is a cost you pay on purpose. The day a managed service closes, your roadmap becomes theirs.

Q14. True or False: S3 Vectors and OpenSearch Serverless sit on the Memory layer, because both persist data across sessions.

• A) True
• B) False

Show answer

B) False. Both are vector stores on the Knowledge layer. Persisting data is not the same as session memory.

Q15. On the three axes, which end is the safest default for a first production launch?

• A) whichever end fits the job in front of you, not the fanciest
• B) always the richer, more capable end, since it tends to scale further later
• C) always managed, one agent, stateless
• D) always build-your-own, many agents, stateful

Show answer

A) Same capability, different cost, control, and effort at each end. The job picks the end, not a preference for power.