Review Mode Set 4 - Claude Certified Architect Foundations

1. Question

Category: CCAF – Agentic Architecture & Orchestration

You are building a financial reconciliation agent using the Claude Agent SDK. The agent processes bank transactions by calling match_transaction, then verify_balance, then post_reconciliation. The agentic loop currently uses a counter that terminates after 15 iterations as a safety measure. During testing, complex reconciliation batches with many exceptions require 18–20 iterations to complete, causing the agent to stop prematurely with unreconciled items. Simple batches finish in 3–4 iterations but still run through the remaining iterations, producing unnecessary “checking for more items” messages.

What is the correct approach to fix both issues?

Increase the iteration cap from 15 to 25 to accommodate complex batches while keeping the safety limit.
Replace the iteration cap with stop_reason-based control flow: continue looping when stop_reason is 'tool_use' and terminate when stop_reason is 'end_turn', using the iteration cap only as a secondary safety net rather than the primary stopping mechanism.
Parse the agent’s text output for phrases like 'reconciliation complete' or 'all items processed' to determine when to stop the loop.
Implement a separate monitoring agent that watches the reconciliation agent's progress and signals termination when no new transactions are being processed.

2. Question

Category: CCAF – Agentic Architecture & Orchestration

You are building a multi-agent system for competitive intelligence. Your coordinator agent decomposes research queries and delegates to subagents. When a user asks, “Compare our product pricing against the top 3 competitors,” the coordinator creates three separate tasks: “Research Competitor A pricing,” “Research Competitor B pricing,” and “Research Competitor C pricing.” Each subagent successfully returns detailed pricing data. However, the final report lacks a comparative analysis and reads as three independent pricing summaries stapled together.

What is the most likely architectural issue?

The coordinator needs an additional synthesis subagent that takes all three pricing reports and produces a comparative analysis.
The coordinator should invoke the three research tasks sequentially so each subagent can build on the previous one's findings.
The subagents need access to each other's findings during their research phase so they can identify comparison points.
The coordinator's prompts specify step-by-step procedural instructions for data gathering rather than research goals and quality criteria that would enable the report agent to produce integrated analysis.

3. Question

Category: CCAF – Agentic Architecture & Orchestration

Your coordinator agent manages a research pipeline where the synthesis subagent has produced a draft report. The coordinator evaluates the report and identifies three gaps: missing data on Asian markets, no discussion of regulatory impacts, and insufficient sourcing for the cost projections. Rather than accepting the incomplete report, the coordinator needs to improve it.

What orchestration pattern should it follow?

Send the report back to the synthesis subagent with feedback requesting it fill in the gaps from its existing knowledge.
Use an iterative refinement loop that sends targeted gap queries to the search and analysis subagents, then re-runs synthesis with the new findings.
Accept the current report and append a disclaimer listing the identified coverage gaps.
Spawn three parallel subagents, one for each gap, and merge their outputs into the existing report without re-running synthesis.

4. Question

Category: CCAF – Agentic Architecture & Orchestration

You are configuring subagents for your multi-agent research system using the Claude Agent SDK. The coordinator needs to spawn subagents, but during testing, the coordinator generates text describing what it would delegate rather than actually invoking the Task tool. Upon reviewing the coordinator’s AgentDefinition, you discover that allowedTools is set to ["web_search", "read_document", "write_report"].

What is the issue?

The coordinator's system prompt doesn't include instructions to delegate tasks to subagents.
The subagent AgentDefinitions haven’t been registered with the system, so the coordinator doesn’t know they exist.
The coordinator’s allowedTools configuration must include the Agent tool to invoke subagents. Without this tool, the coordinator cannot spawn subagent processes.
The coordinator needs tool_choice set to 'any' to force it to use tools rather than generating text.

5. Question

Category: CCAF – Agentic Architecture & Orchestration

Your customer support agent handles a message: “I need to cancel my subscription, get a refund for this month since the service was down for a week, and update my email address to sarah.new@tutorialsdojo.com.” The agent currently processes one concern at a time sequentially, first canceling the subscription, then investigating the refund, then updating the email. However, the refund investigation takes 30 seconds because it requires checking service uptime logs, and the email update is instant.

What is the most efficient handling approach?

Process all three concerns sequentially in the order the customer listed them to maintain predictability.
Prioritize the most time-sensitive concern, which is cancellation, and handle the other two only after confirming that the cancellation is complete.
Ask the customer to submit three separate support tickets, one for each concern.
Decompose the request into distinct items, investigate each concern using shared context, and handle independent items in parallel where possible before synthesizing a unified resolution.

6. Question

Category: CCAF – Context Management & Reliability

You are building a handoff workflow for your customer support agent. When the agent determines it needs to escalate to a human supervisor, it currently sends the message: “I’m transferring you to a supervisor now.” The human supervisor then has to spend 5-10 minutes reading through the full conversation transcript to understand the situation before the supervisor can help the customer.

How should you improve the escalation process?

Provide a structured handoff summary with the customer ID, issue, actions taken, refund details, and recommended next steps.
Include a link to the conversation transcript in the handoff message so the supervisor can review it quickly.
Transfer the customer and let the supervisor ask the customer to re-explain their issue, since the customer can provide the most accurate account.
Have the agent provide a one-sentence summary: "Customer is requesting a refund for a defective product."

7. Question

Category: CCAF – Agentic Architecture & Orchestration

Your multi-agent research system has a coordinator that delegates to a web search agent, a document analysis agent, and a synthesis agent. You notice that the web search agent and document analysis agent sometimes produce overlapping findings because both agents research the same subtopics and return near-duplicate information. This wastes tokens and confuses the synthesis agent, which sometimes presents the same finding twice with different source attributions.

How should you address this?

Add deduplication logic in the coordinator that compares subagent outputs and removes duplicates before passing to synthesis.
Run the subagents sequentially so the second agent can see what the first already covered.
Have the subagents communicate directly to coordinate their research and avoid overlap.
Partition the research scope across subagents to minimize duplication by assigning distinct subtopics or source types to each agent.