Product design question

Claude Code can spawn hundreds of parallel subagents in one session. What risks would you design for?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Tests systems thinking about scale risk: what breaks when an agentic product multiplies its own actions hundreds of times over.

Clarify the scenario: hundreds of subagents running in parallel inside one Claude Code session, each able to read, write, and execute code.
Identify the resource risks: runaway cost from token usage, hitting rate limits, and exhausting local compute or file system resources.
Identify the correctness risks: subagents conflicting on the same files, duplicating work, or compounding a shared bad assumption across all of them.
Identify the safety risk: a single subagent taking a destructive action, like deleting files or pushing code, multiplied by hundreds of parallel instances.
Propose design mitigations: a shared lock on files being edited, a global budget cap, and a required human approval gate before any destructive or external action.
Confirm with the interviewer whether the concern is cost and correctness, safety, or both, since that changes which guardrail to prioritize.

Names concrete failure modes, like two subagents editing the same file and silently overwriting each other's work, not just abstract risk.
Proposes a hard budget and rate ceiling per session so a runaway loop cannot spend unbounded tokens or API calls.
Suggests observability: a live view of what each subagent is doing, so a user can intervene before hundreds of actions compound.
Distinguishes reversible actions, like reading a file, from irreversible ones, like a git push or a delete, and gates only the latter.

Only discussing cost risk and missing correctness risk from parallel subagents stepping on each other.
Proposing a blanket approval requirement for every action, which would defeat the point of parallelism.
Not addressing observability, leaving the user unable to see what hundreds of subagents are actually doing.

Chapters of the AI PM course that teach what this question tests.