Agentic AI Interview Questions
Dashboard
0%
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →

Q41HardScenario

Your team built a 6-agent system. It costs 10x more than the old single agent, with no quality improvement. What do you do?

30-second answerSay your answer out loud first, then reveal.

Diagnosis

  1. Per-agent breakdown: tokens, latency, and a quality contribution per agent from traces. Which agents actually change the outcome?
  2. Ablation study: remove or merge agents one at a time and measure on the eval set. Often 2–3 agents are doing what one could.
  3. Hand-off analysis: is the full conversation history copied to every agent? That multiplies input tokens.
  4. Coordination overhead: supervisor round-trips after every small step; agents "discussing" with each other.
  5. Duplicated work: two agents searching the same thing because the briefs overlap.

Typical fixes

  • Collapse into a single agent with a well-designed toolset and plan/to-do list.
  • Agents-as-tools: keep one main agent and call specialised sub-agents only for heavy, isolatable work.
  • Pass structured briefs instead of full transcripts.
  • Cheaper models for workers, the strong model only for the orchestrator or synthesis.
  • Effort scaling: the router decides whether a query needs the multi-agent path at all.
What to say about multi-agent value: multi-agent pays off for breadth-first, parallelisable tasks (research across many independent sources) and where context isolation matters. It rarely helps for sequential, tightly coupled tasks. Matching the architecture to the task's structure is the key judgement.
PreviousAn email-assistant agent was tricked by a malicious email into forwarding confidential documents to an attacker. Explain the attack and design defences.

Little by little, you're building something great.