AnthropicClaude Certified Architect – Professional2026-101.3 Select appropriate architectural patterns (workflow, agentic, augmented LLM)
Defend the simplest acceptable production pattern
Completed by the knowledge check
Learning objective
Compare candidate patterns on autonomy, blast radius, auditability, latency, cost, and fallback behavior, choosing the simplest that meets requirements.
Core idea
The simplest acceptable pattern is the one that satisfies the result contract with the least necessary execution machinery. It need not be one model call, and it need not be the cheapest successful demonstration. Mandatory inspection, permitted effects, deadlines, and incomplete-work handling are part of the requirements.
Anthropic recommends beginning with simple prompts and adding multistep execution only when simpler solutions fall short. Turn that principle into a pattern-selection record: compare candidates on the same workload, reject those that violate requirements, and explain the operating costs you accept for the remaining choice.
The six dimensions are a comparison lens, not an automatic score. A lower token estimate cannot compensate for an unacceptable external action, and better average quality cannot compensate for a missing approval gate. Apply hard constraints before ranking the acceptable candidates.
How it works
Compare autonomy and blast radius separately
Autonomy concerns how much progression the model directs; see 1.3.1. Blast radius is the scope of harm a wrong result or action can cause. They are related, but permissions and delivery boundaries determine whether a mistake reaches an external system.
Compare the effect surface, not only the pattern name. A read-only investigator that produces an internal draft can have more path discretion and less operational exposure than a fixed workflow that automatically publishes an answer. Record which candidate can read sensitive records, mutate state, or deliver output without review.
For an SDK candidate, the enforcement point is the configured tool environment and permission policy. A PreToolUse hook can deny a matching operation through hookSpecificOutput.permissionDecision: "deny"; see 1.2.2 for the complete deny fields. Compare candidates with that policy included, rather than comparing a guarded workflow against an unrestricted prototype agent.
Compare auditability through actual evidence
A predefined workflow gives reviewers named stages and transitions. That makes expected-path inspection easier, but does not itself save the evidence. An autonomous path requires the reviewer to reconstruct what observations led to the next action.
Anthropic warns that framework abstractions can obscure prompts and responses. Require access to the inputs, intermediate outputs, and decisions your acceptance review needs before accepting an abstraction. A diagram of intended stages is not a record of a completed execution.
The TypeScript Agent SDK’s SDKResultMessage provides session_id, duration_ms, num_turns, total_cost_usd, and a result subtype for run-level accounting. Its permission_denials is the reference’s authoritative denial record; each SDKPermissionDenial carries tool_name, tool_use_id, and tool_input. Do not assume this records denials made by a PreToolUse hook.
For per-action evidence, enable the SDK’s OpenTelemetry export with CLAUDE_CODE_ENABLE_TELEMETRY=1 and an exporter. OTEL_LOGS_EXPORTER=otlp exports log events, including a record for each tool result. Traces add claude_code.tool spans per invocation and require OTEL_TRACES_EXPORTER=otlp plus CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1; tracing is beta, and span names and attributes can change. Export errors fail silently by default, so verify incoming records at the collector before relying on them for an audit.
Join this execution evidence with the accepted artifact and unresolved checks. A harness subtype: "success" still needs business acceptance.
Compare latency and cost over the whole path
Count the work from intake to an accepted outcome. A single prepared-context call has fewer external stages than a chain; a chain introduces ordered calls and gates. Independent calls can run concurrently, while an agent’s number of tool rounds varies with its investigation. See 1.3.2 for the task shapes that justify those calls.
Include retrieval, joins, evaluation, and fallback in the elapsed-time measurement. A fast draft followed by an expensive repair loop is not a fast accepted answer. Measure deadline misses and unresolved cases alongside accepted-case timing so failed executions do not disappear from the comparison.
For SDK experiments, use duration_ms and num_turns to characterize the harness, and total_cost_usd for estimated run cost. The TypeScript reference explicitly describes the latter as an estimate, not a billing statement. Include application work outside the harness separately, and document session accounting if you resume runs rather than comparing unlike totals.
An agent’s configured maxBudgetUsd is an estimated stopping threshold with headroom requirements, not an exact financial reservation. See 1.3.3 for the overshoot boundary. Reserve operating allowance for incomplete-work handling.
Compare fallback as a terminal behavior
For each candidate, specify what happens when evidence is missing, a gate rejects output, a required call fails, or the run exhausts its limit. The fallback artifact should state the unresolved case, preserve useful observations, and name the receiving owner.
Fallback can be staff triage, a clarification request, or another evaluated processing path. These are application decisions, not an automatic architectural property of workflows or agents.
An SDK result with error_max_turns, error_max_budget_usd, or error_during_execution enters incomplete-work handling. Do not infer completion from the final draft. Do not enlarge limits automatically until a run appears successful; that removes the operating constraint used to justify the candidate.
Evaluate fallback demand against the team’s capacity. If the simpler candidate transfers too many unresolved cases to an understaffed queue, it fails an operating requirement even if its accepted answers are excellent. Conversely, autonomy that reduces that queue only slightly may not justify its extra runtime and investigation risk.
Key distinctions and trade-offs
Use the following matrix as a starting hypothesis to test, rather than a guaranteed ranking:
| Dimension | Prepared-context answering step | Prescribed workflow | Agentic investigation |
|---|---|---|---|
| Autonomy | Bounded answering task | Known transitions, with scoped model decisions | Model directs the investigative path |
| Blast radius | Depends on how the answer is used | Depends on tool effects and mandatory gates | Depends on permitted effects and exit boundaries |
| Auditability | Supplied evidence and answer | Stage outputs and transition decisions | Observations, selected actions, policy decisions, and terminal result |
| Latency | Few external stages | Sequential dependencies and joins | Variable rounds and exploration |
| Cost | One answering invocation plus preparation | Calls purchased by the selected path | Variable calls, tool work, and unsuccessful exploration |
| Fallback | Missing evidence or rejected answer | Failed stage, gate, or unresolved category | Blocker, rejected result, execution failure, or limit exhaustion |
A hybrid can preserve a simple routine path and pay for investigation on exceptions. It also adds a routing and boundary contract that the team must operate. Compare that total responsibility with the workload benefit; combining patterns is not free simplicity.
Architecture / decision scenario
A healthcare provider prepares internal administrative policy explanations. The hypothetical requirements are a ten-second routine-answer target, reviewable source evidence, no automatic changes to patient records, and an exception queue that two staff members can handle.
The team compares a prepared-context call, a prescribed retrieval-and-review workflow, and an autonomous investigator on the same routine and unusual cases. It measures accepted explanations, unsupported claims, elapsed time, estimated cost, and staff fallback demand. The scenario does not assume any vendor benchmark result.
Suppose the prepared-context call meets the routine target and quality bar, but unusual requests overwhelm the queue. The workflow improves inspection while still missing evidence whose location depends on the case. The investigator resolves enough exceptions to fit staff capacity, but misses the routine target when used everywhere.
The architect selects routine answering plus a bounded investigation branch for exceptions. The permission boundary prevents record mutations; investigation outputs require review before use. The record compares all six dimensions: scoped autonomy, contained effects, evidence traceability, separate latency targets, estimated spend with headroom, and a staffed incomplete-work path.
The team rejects an investigator for routine requests because it adds latency and cost without useful gain, and rejects single-call-only delivery because fallback demand exceeds staffing. The simplest acceptable choice is the hybrid under these measured assumptions, with a review trigger if its exception demand or acceptance results change.
Common exam traps
- Ranking candidates by token cost before applying hard requirements.
- Equating low autonomy with a small blast radius.
- Treating a stage diagram or result summary as complete audit evidence.
- Measuring draft latency while excluding validation and fallback.
- Treating a spend estimate as a billing statement.
- Choosing a simple primary path whose exception queue cannot operate.
- Replacing incomplete-work handling with repeated increases to agent limits.
What to remember
Compare the permitted effects and the full accepted-outcome path. Apply hard constraints first, then choose the simplest candidate that meets quality, timing, spend, evidence, and fallback capacity requirements. Record the rejected alternatives and the conditions that would change the decision.
Knowledge check
A bank compares three candidates for drafting customer replies. A requirement states that no reply reaches a customer without staff review. On the same workload, the single call is cheapest and sends replies automatically; the autonomous investigator scores highest on quality and also sends automatically; the prescribed workflow routes every draft to staff review and meets the quality bar and the deadline at moderate cost. Which candidate should the selection record choose?
Two candidates handle billing disputes. A read-only investigator chooses its own lookups and writes an internal draft that staff review. A fixed workflow follows prescribed steps and sends its answer directly to the customer. Which candidate has the smaller blast radius?
An insurer answers internal policy questions with a ten-second routine target and an exception queue that two staff members can handle. Tested on the same cases, a single prepared-context call meets the routine target but sends far more unusual requests to the queue than staff can handle. An investigator resolves enough unusual requests to fit staff capacity but misses the routine target when used for every request. Intake already distinguishes routine from unusual requests, and the measured timings include branch selection. Which design is the simplest that meets the requirements?
An Agent SDK run for a pattern experiment ends with a result subtype of error_max_turns. Its last assistant message contains a plausible draft. An engineer proposes doubling maxTurns and rerunning whenever this happens until the run succeeds. What should the team do with this run?
maxTurns and rerun until the result reports successmaxBudgetUsd raised instead of maxTurnsReferences
- Anthropic, Building effective agents
- Anthropic, TypeScript SDK reference
- Anthropic, Agent loop
- Anthropic, Control execution with hooks
- Anthropic, Configure monitoring and observability