Skip to lesson

AnthropicClaude Certified Architect – Professional2026-106.2 Communicate architectural decisions and trade-offs

Defend the choice against credible alternatives

Completed by the knowledge check

Learning objective

Defend decisions against credible alternatives with evidence from evaluations and prototypes.

Core idea

Defending a decision means showing why it follows from requirements and observations, including observations that favor a different option. A persuasive demo cannot settle an operating-cost dispute, and repeating a preferred design’s benchmark cannot answer a challenge about a missing capability.

Produce a comparison dossier: the disputed claim, viable alternatives, evaluation results, prototype observations, unresolved questions, and the condition that changes the recommendation. NIST AI RMF MEASURE 2.1 calls for documented test sets, metrics, and evaluation tools; MEASURE 2.3 calls for assessment under conditions similar to deployment. Use those identifiers to make the argument inspectable rather than treating “we prototyped it” as sufficient support.

How it works

Turn the objection into a testable claim

First identify the disagreement. “The agent would handle exceptions better” is a capability claim. “Flexibility is worth a higher bill” is a priority judgment. “We should permit automatic changes” is a question of authority and risk tolerance. More evaluation can inform all three, but it cannot decide organizational priorities or authorize risk acceptance by itself.

Record the strongest version of the alternative, including its intended controls. Do not compare a guarded workflow with an unconstrained agent if the challenger proposes an agent with the same authorization checks. If the alternative cannot meet a mandatory requirement, name the missing capability or dependency and let its proponent respond before excluding it.

The challenge should produce a concrete evaluation question: on the disputed exception slice, does the agent increase accepted outcomes enough to justify its measured latency, cost, and implementation burden? This keeps the response connected to the decision criteria from 6.2.1 rather than becoming a debate about whether agents are inherently better.

Identify what each artifact proves

An evaluation estimates performance on defined tasks with stated graders. A prototype exercises a proposed implementation and can expose integration difficulty, missing permissions, or an unexpected sequence of calls. A load test addresses behavior under its specified traffic conditions. None automatically establishes the others.

Anthropic’s evaluation guidance distinguishes the transcript from the outcome. A transcript records what the agent did; the outcome is the final state of the environment. To defend a claim that a workflow completes a business operation, bring verified outcomes as well as plausible explanations. A model saying it updated a record is weaker support than the record actually having the required value.

For the comparison dossier, preserve enough provenance to connect each observation to the tested configuration and acceptance rule. For a Messages API experiment, record model, the prompt in system and messages, available tools, and the output limit max_tokens, alongside any application execution limits. The purpose here is not to recreate the evaluation methods in subsection 4.2, but to ensure a reviewer can tell whether the challenged architecture is the one that produced the result.

Compare complete operating proposals

Use the same task population and success definition, and disclose material differences between configurations. If one option uses retries or expert selection and the other receives one attempt, report that operating difference with its cost and latency. A trial budget is part of the proposal, not a detail to remove before the executive comparison.

Anthropic recommends isolated trial environments because leftover state can inflate results or introduce correlated failures. Ask whether the prototype comparison used comparable clean starting conditions. If the preferred option inherited a solved task’s state, its result cannot support the recommendation until the comparison is repaired.

Show both whole-workload and disputed-slice results. Keep the row on which the alternative wins. If the evaluation does not include the disputed case, the honest answer is a bounded follow-up experiment, not an extrapolation from unrelated scores. Anthropic recommends calibrating model-based graders against human experts; a contested quality difference may require adjudication before it becomes a reason to spend more.

Say what would change your answer

Conclude with the requirement that decides the comparison, the remaining uncertainty, and a measurable reconsideration condition. NIST MANAGE 1.1 connects assessment to whether development or deployment should proceed, while MANAGE 1.3 identifies response options for risks. The authorized decision maker applies the organization’s priorities and risk tolerance to the observations.

If the challenger identifies a real omission, revise the recommendation or limit its scope. If the comparison remains decisive, retain the decision and record the dissent with the condition for reopening it. A trigger should name a changed assumption or observable result, not merely a calendar date that guarantees another inconclusive meeting.

Key distinctions and trade-offs

Credible alternative versus straw man: A credible alternative can plausibly satisfy the business need and includes reasonable controls. Rejecting a deliberately unsafe version does not defend your design against the version another team actually proposes.

Capability versus adoption case: Better exception handling may justify an agent for a narrow queue without justifying it for the entire workload. Anthropic’s guidance to add complexity when it demonstrably improves outcomes supports scoped adoption rather than a universal architecture preference.

More observations versus a different decision: A capability claim needs a suitable comparison. A disagreement about acceptable cost needs a decision owner. A risk-acceptance disagreement needs the authorized risk owner. An additional prototype cannot resolve unclear authority.

Each disagreement goes to what can settle it: a capability claim to a comparison that revises or retains the decision, an acceptable-cost disagreement to a decision owner, and a risk-acceptance disagreement to the authorized risk owner.

Each disagreement goes to what can settle it, a capability claim to a comparison that revises or retains the decision, acceptable cost to a decision owner, and risk acceptance to the authorized risk ownerStarts at identifying the disagreement and branches three ways. A capability claim goes to a suitable comparison; a real omission revises the recommendation or limits its scope, and a comparison that remains decisive retains the decision and records the dissent. A disagreement about acceptable cost goes to a decision owner, and a risk-acceptance disagreement to the authorized risk owner, not to another prototype.capability claimacceptable costrisk acceptancereal omissionremains decisiveIdentify thedisagreementSuitablecomparisonDecision ownerAuthorized riskownerRevise or limit itsscopeRetain the decisionRecord the dissent

Comparison costs engineering time. Run the smallest experiment that can settle the deciding uncertainty, while preserving production-relevant controls and success criteria. Name what remains untested instead of inflating a focused prototype into a full operational qualification.

Architecture / decision scenario

A financial-services operations team processes 2,000 internal reconciliation cases daily. Ninety percent follow stable rules; 10% require investigation. The team has four engineers and a $1.50 ceiling per accepted case, including review. The architecture proposal uses a workflow for routine cases and manual investigation for exceptions. A senior engineer credibly proposes a tool-using agent with the same write-approval controls for both groups.

In a hypothetical comparison on 800 cases matching that 90/10 mix, the workflow passes 691 of 720 routine cases, while the agent passes 698 of 720. On the 80 exceptions, manual investigation passes 36 and the agent-assisted process passes 64. Estimated cost per accepted case is $0.60 for the workflow proposal and $1.80 for the agent proposal. A sandbox prototype confirms that the agent can complete the exception investigation integration; it has not tested production peak load.

Do not defend the workflow by quoting only the routine-case score. The exception advantage is material, and the challenger has preserved the authorization requirement. Also do not switch every case to the agent: the measured proposal exceeds the cost ceiling, and the prototype does not establish production capacity.

Recommend retaining the routine workflow and funding a bounded pilot of the hybrid: a workflow for routine cases and agent-assisted investigation for exceptions. Measure its accepted-case cost against the $1.50 ceiling under representative load. The dossier includes outcome checks from the sandbox, the exception results, cost assumptions, and the untested load condition. It identifies the quality improvement as observed, integration feasibility as prototyped, and production economics of the hybrid as unresolved.

The operations director owns the quality-cost priority, and the engineering lead owns the bounded hybrid pilot. Reopen the exception decision when the hybrid demonstrates its agreed acceptance rate and total accepted-case cost within the ceiling under representative load. This response defends the routine choice, concedes the alternative’s demonstrated advantage, and directs the next investment toward the uncertainty that matters.

Common exam traps

  • Dismissing a stakeholder’s viable alternative without implementing its proposed controls.
  • Treating a transcript’s success claim as a verified business outcome.
  • Comparing unequal retry budgets while hiding cost and latency differences.
  • Using a sandbox integration demo as proof of production capacity.
  • Defending an aggregate score against an objection about a particular task slice.
  • Running more experiments when the unresolved issue is priority or risk authority.

What to remember

Answer the actual objection with the artifact that can settle it. Compare credible controlled alternatives on matching tasks, verified outcomes, and complete operating costs. Preserve the alternative’s strongest result, distinguish prototype feasibility from deployment performance, and state the observation that would change your recommendation.

Knowledge check

References