Practice — Agentic Architecture & Orchestration
20 questions — none attempted yet.
A team builds a support bot: it reads a customer's message, decides whether it needs to look something up, calls a tool if so, reads the result, and decides whether to answer or look up something else — repeating until it has enough information to reply.
Which single design element makes this an *agent* rather than a single-shot completion?
An engineer implements an "agent": it calls Claude once with a list of tools, executes at most one resulting tool call if present, and returns whatever it has to the user — regardless of whether the model's response indicated it wanted to keep going.
What is actually missing from this implementation?
A coding agent edits a file, runs the test suite, reads any failures, edits again, and repeats.
Which condition should end this loop?
A company needs to convert a fixed-format CSV export into a fixed-format JSON file. The mapping never varies and involves no judgment calls.
What architecture fits this task?
A hiring pipeline auto-rejects any candidate below a fixed GPA cutoff with no exceptions. Candidates above the cutoff go through a nuanced fit review where judgment genuinely varies case by case.
How should this be architected?
An expense-approval system must: read a submitted receipt, check it against fixed policy limits, flag anomalies for finance review, and auto-approve clear cases.
Which architecture fits these requirements?
A research assistant needs to answer one narrow factual question by searching, reading a couple of sources, and synthesizing a short answer.
Should this be a single agent or a multi-agent system?
A research assistant needs to independently investigate five unrelated competitor companies and produce one paragraph per company. None of the five investigations depends on any other.
What's the best architecture?
A single Claude agent is given 40 tools spanning HR, finance, IT, and legal systems, and is expected to handle any employee question end to end, alone.
What's the issue, and the best fix?
A team debates what "subagent" means in an orchestration design.
Which definition is accurate?
An orchestrator agent delegates "summarize this 200-page contract" to a subagent, then continues the conversation with the end user.
What should the orchestrator receive back from the subagent?
A team is deciding when it's worth running tool calls or subagents in parallel.
When does parallelization actually help?
An agent needs order status from System A and a shipping ETA from System B to answer a customer's question. Neither lookup depends on the other's result.
What's the best approach?
A team is deciding where to add a human-in-the-loop step to an agent.
What's the primary reason to add a human-in-the-loop gate?
A refund agent is required to "escalate unusual cases to a human" rather than deciding on its own.
Which implementation actually satisfies that requirement?
A team is defining stopping conditions for a new agent loop.
Which of these is a well-formed stopping condition?
An agent researching a topic keeps calling a search tool, because each new search "might turn up something more."
What's missing from this design?
An agent's only stopping condition is a hard cap of 50 turns.
What's the risk, and what's the fix?
A team is weighing an agent-based design against a deterministic workflow for a new feature.
What's the main reliability trade-off of choosing an agent over a deterministic workflow?
A team is choosing between a single-agent and multi-agent design for a task that is strictly sequential — each step depends on the last. They pick multi-agent, reasoning that "more agents double-check each other" for extra reliability.
Evaluate this reasoning.