Architect multi-agent solutions
Architecture questions test whether you can decompose a business process into agents with clear responsibilities and choose the right coordination pattern before any code exists. Know the canonical topologies: orchestrator (supervisor) with specialist workers, sequential pipeline, hierarchical (orchestrator of orchestrators), and peer/blackboard collaboration. Match the pattern to the work: deterministic multi-step workflows favor sequential or orchestrator patterns; open-ended research favors routing with a planner. Understand that an agent is a bounded role with instructions, tools, and a model - not a microservice and not a single prompt. The key design decisions are role boundaries, the handoff contract (what payload passes between agents), shared state versus message passing, and where human approval gates belong.
Traps to expect: questions that push you toward a single monolithic agent when the scenario emphasizes independent specializations or separate security scopes; scenarios where adding more agents increases latency and cost without improving accuracy; and confusing the orchestrator's routing role with a worker's domain reasoning. Another common trap is assuming agents share memory automatically - they do not unless you design a shared state store or explicit context passing. Watch for scenarios requiring deterministic control flow, where a rigid pipeline beats autonomous delegation.
Rule of thumb: if two responsibilities need different tools, different data access permissions, or different evaluation criteria, they belong in different agents; if they always execute together in a fixed order with no branching, keep them in one agent.
Worked example: a claims intake scenario needs document extraction, policy lookup, and a fraud decision with human review. Architect as an orchestrator that routes to an extraction agent, a policy agent, and a fraud agent, with a human approval gate before any payout action - not one agent with all tools, because the fraud agent needs a narrower data scope and its own evaluation set.