Rubric
Contents — domains, guide and mocks

Communicating decisions and trade-offs

CCAR-P 6.210 min read · checked 21 September 2026

Task statementCommunicate architectural decisions and trade-offs

How a trade-off conversation runs

  1. Restate the goalthe agreed success criteria
  2. Show 2–3 optionseach with cost, speed, risk
  3. Recommend oneand what it gives up
  4. Owner decidesrecorded as a decision record
The recommendation comes after the options, and the decision is recorded — so nobody later asks “why did we do it this way?”.

The trade-offs Claude architectures actually involve

Most decisions on a Claude engagement pull on the same few levers: quality of output, latency, cost per task, operational complexity, and risk (what happens when the system is wrong). Anthropic’s Building effective agents article states the core one plainly — agentic systems often trade higher latency and cost for better task performance, so you should add that complexity only when it pays for itself. Workflows (predefined code paths) are predictable and cheaper; agents (the model directing its own steps) handle open-ended problems but are harder to predict and test. That single sentence is the backbone of many stakeholder conversations.

DecisionOption AOption BWhat the listener gives up
Workflow or agentFixed steps, predictableModel chooses steps, flexibleA: handles fewer unusual cases. B: more cost, latency and testing effort
Model sizeLargest modelSmaller, faster modelA: higher cost and latency. B: may miss the accuracy target on hard cases
Live or batchAnswer in secondsOvernight batchA: pay for real-time capacity. B: users wait for results
Human reviewReview every outputReview only flagged onesA: slower and staff-heavy. B: some errors reach customers
Build or buyCustom integrationOff-the-shelf connectorA: time and maintenance. B: less control over behaviour

Same decision, different audiences

The CCAR-P guide page on this site advises rehearsing the explanation, not just the decision: how you would justify a choice to a sceptical engineering lead, and separately to the person paying. The facts do not change between audiences; the vocabulary and the level of detail do. An executive needs the business consequence and the decision you need from them. An engineering lead needs the mechanism, the failure modes and what they will have to operate. A compliance officer needs where data goes and where a human decides.

One decision, two framings

To the engineering leadtext

Proposal: fixed 3-step workflow
(classify -> retrieve -> draft)
instead of an autonomous agent.
- Deterministic path, easy to
  unit-test each step.
- Unusual tickets (~8% in the
  sample) route to a human queue.
- Revisit if the human queue
  exceeds 15% for 4 weeks.

To the COOtext

Recommendation: start with a
guided workflow, not a free-
roaming agent.
- Lower running cost, faster
  answers, easier to audit.
- About 1 in 12 tickets still
  goes to your team.
- Decision needed: approve the
  4-week pilot by 3 October.
Both versions say the same thing. The executive version leads with money, risk and the ask; the engineering version leads with mechanism and operations.

Recording the decision

A decision explained only in a meeting is forgotten in a quarter. Michael Nygard’s Architecture Decision Record (ADR) format is the widely used remedy: a short document per significant decision with a title, the context (the forces at play, written neutrally), the decision itself (“We will…”), its status, and the consequences — positive, negative and neutral. Nygard’s argument is that people joining later otherwise either accept a decision blindly or reverse it blindly, because they cannot see why it was made. Writing down the negative consequences is the part teams skip, and it is exactly the trade-off stakeholders agreed to.

A trade-off ADR, kept to one pagetext
ADR-007: Route product questions by complexity

Status: Accepted (2026-09-18)

Context: Answers must be >=90% correct on the 400-question
set, first token < 2 s, within the monthly budget. The
largest model alone meets accuracy but exceeds the budget.

Decision: We will answer with a smaller model by default
and route questions classified as complex to a larger model.

Consequences:
+ Meets accuracy and budget on the test set.
- Adds a classifier to build, test and monitor.
- A misrouted complex question gets a weaker answer.
= Revisit if routed-question accuracy falls below target.

How to write and maintain ADRs as part of the wider documentation set is covered in 6.4; here the point is that the record captures the trade-off in the words the stakeholders approved.

Being honest about uncertainty and risk

Language models are probabilistic, and a professional architect says so without alarming anyone. Replace “it will be accurate” with “on the 300-case test set it was right 96% of the time; the errors were mostly on handwritten forms, which is why a person checks those”. State what you measured, what you have not yet measured, and what you will do if production differs. Where a platform commitment matters, quote the provider’s own terms rather than your impression — for example, the Claude API docs describe a Priority Tier with a 99.5% uptime target for organisations with existing capacity commitments, while the standard tier is best-effort. SLAs themselves are covered in 6.3.

Checking a decision summary before you send it

  • Passes: States the goal in agreed success criteria
  • Passes: Shows at least two real alternatives
  • Missing: Names what the chosen option gives uponly benefits listed
  • Check: Evidence is measured, not assertedcost figure is an estimate
  • Fails: Says who decides, and by whenno owner named
  • Passes: Written for this audience’s vocabulary
A draft summary for a bank’s risk committee, reviewed against what a good trade-off write-up needs.

Traps the wrong answers are built from

Tempting but wrongDo this instead
Presenting a single design as the only possible answer.Show two or three real options with their costs and risks, then recommend one.
Listing only the benefits of the recommended option.State what it gives up — cost, latency, residual error, operational load — in plain words.
Giving executives the same technical deep-dive as engineers.Keep the facts identical but lead with business consequence and the decision needed.
Backing a choice with impressions (“the big model is better”).Measure the options against the agreed test set and report the numbers.
Leaving the decision in meeting notes or someone’s memory.Record it as an ADR with context, decision, status and consequences.

You should now be able to

  • Identify the main trade-off levers in a Claude architecture: quality, latency, cost, complexity and risk.
  • Present two or three options with consequences and a recommendation tied to agreed success criteria.
  • Reframe the same decision for executive, engineering and compliance audiences without changing the facts.
  • Capture a decision and its negative consequences in an architecture decision record.
  • Communicate model uncertainty with measured results rather than guarantees.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    You recommend a fixed workflow rather than an autonomous agent for an insurer’s claims triage. The CIO asks why, since “agents are the future”.

    Which response best communicates the decision?

    1. AAgents are unreliable, so we avoid them in regulated industries like insurance.
    2. BIt met the agreed accuracy target; an agent adds cost and latency we can’t yet justify.
    3. CWe can build an agent instead if you prefer; both approaches would work fine.
    4. DHere is a detailed diagram of the workflow’s prompt chain and retry logic.
    Show answer and reasoning
    1. AIncorrect. A blanket claim with no evidence or link to this client’s goals sounds like opinion and invites the CIO to overrule it.
    2. BCorrect. It ties the choice to the agreed criteria, states the trade-off Anthropic’s guidance describes, and leaves room to revisit.
    3. CIncorrect. Deferring to preference abandons the recommendation and hides the real differences in cost, latency and testability.
    4. DIncorrect. Mechanism detail answers an engineer’s question, not the CIO’s question about why this option serves the business.

  2. Question 2

    A hospital’s board must approve an architecture for drafting discharge letters. You have a 40-page technical design.

    Which two things should the summary for the board contain? (Select 2.)

    1. AThe decision needed from the board and the deadline for it.
    2. BOnly the benefits, so the board is not unduly worried about AI.
    3. CThe full prompt text and retry settings for each pipeline step.
    4. DThe residual risk of the chosen design and how it is controlled.
    5. EA statement that the model will always produce correct letters.
    Show answer and reasoning
    1. ACorrect. Senior readers need the ask up front; without it the paper informs but produces no decision.
    2. BIncorrect. Hiding downsides undermines trust and leaves the board approving a risk it was never shown.
    3. CIncorrect. Accurate but irrelevant to the board’s decision; it belongs in implementation documentation.
    4. DCorrect. Stating what the option gives up — and the human review that controls it — is the honest core of a trade-off summary.
    5. EIncorrect. No probabilistic system can promise this; the claim sets an expectation that will be broken.

  3. Question 3

    Eighteen months after launch, a new engineering manager at a BPO proposes replacing the complexity-based model router with the largest model everywhere, “to keep it simple”.

    What would most help the team evaluate this proposal?

    1. AAsking the original architect to explain the decision from memory.
    2. BApproving the change, since simplicity is a core design principle.
    3. CRejecting the change because the architecture was already signed off.
    4. DThe ADR recording the original context, decision and consequences.
    Show answer and reasoning
    1. AIncorrect. Memories fade and people leave; this is the exact failure a written record prevents.
    2. BIncorrect. Simplicity matters, but reversing a decision without knowing its reasons may break the budget constraint it was made for.
    3. CIncorrect. Decisions can and should be revisited when context changes; blind preservation is as risky as blind reversal.
    4. DCorrect. An ADR shows the budget and accuracy forces behind the router, so the team can check whether they still apply.

Practise every question in Domain 6 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.