01
Domain 01
Agentic Architecture & Orchestration
How agents loop, when to plan, how subagents coordinate, where deterministic enforcement beats prompting.
24 questions
- Q01What is the most effective fix?92% of the time the agent verifies identity before issuing refunds — but the 8% just cost the company three wrong refunds.
- Q02Why is the agent refusing to use them?The agent has the tools it needs and is asking for a customer name instead of using them. Why?
- Q03Is this the right fix?A multi-agent finance research system whose synthesis agent is inventing statistics that don't appear in the sources.
- Q04Why is this the wrong approach?A marketing agent for tasks ranging from "subject line" to "quarterly strategy" — should they share one pipeline?
- Q05What's the most likely cause?By turn 8 of a long support session, the agent has forgotten what the customer stated on turn 1.
- Q26What is the correct fix?An appointment-scheduler agent's loop terminates on the word "done" — and exits prematurely when the model uses it mid-sentence.
- Q27What's the root cause?A research system decomposes "impact of remote work on small businesses" into three flavors of tech — and misses every other industry.
- Q28What's the most likely cause?A subagent ignores brand voice rules the brand manager stated five turns earlier in the coordinator's session.
- Q29How should the coordinator emit the calls so the three subagents run in parallel?Three customer concerns spawned across three separate turns — taking 90 seconds when they could run in parallel.
- Q30What's the cleanest fix?Three carrier APIs return dates in three different formats, and the agent keeps mis-parsing them.
- Q51What's the right pattern?60 turns of careful client diagnosis, and now the team wants to explore two recommendation paths in parallel.
- Q52What's the right pattern?Three concurrent client engagements; every morning the consultant re-pastes context summaries and feels something is being lost.
- Q53What's the architectural fix?By turn 30 of a 200-page contract analysis, the agent says the indemnity clause is in section 4.2 — and three turns later, section 5.1.
- Q54What's the better dispatch pattern?A research coordinator dispatches rigid procedural prompts; subagents read 5 papers even when the search returned 2 strong ones.
- Q55What's the better architectural pattern?Three subagents communicating directly — and a debugging surface where errors live in interstitial space no one owns.
- Q88What's the right framing?When does ReAct beat a fixed pipeline — and when is the fixed pipeline actually right?
- Q89What's the most defensible split?Four candidate tasks for AI agent deployment: thank-you notes, credit approvals, ticket routing, employment terminations.
- Q90When does Debate beat Pipeline or Orchestrator?Pipeline, Debate, or Orchestrator — when does Debate beat the other two?
- Q91What's happening, and how do you fix it?"X grew 12.3% YoY in Q3 2024" → "strong growth recently" → "impressive expansion" — the telephone game in pipelines.
- Q92What's the org-chart-shaped framing?One generalist agent for billing, tech support, and feature requests — or three specialists with a router?
- Q93What's the right diagnosis?Eleven specialist agents. Coordination consumes more dev time than the agents. Latency tripled. Quality up, throughput down.
- Q108What's the most actionable next step?"An AI assistant that helps with writing." Vague. The Canvas wants Problem First, 10x Better, One Core Flow.
- Q109What's the operational distinction?A team has been using "chatbot" and "agent" interchangeably. The lead engineer says these are different things.
- Q110Which UX patterns most directly address user trust in AI products?Users get unexpected output and report frustration. Three UX patterns are specifically for AI products.