Question-driven, not map-driven. If the user's request is "describe this repo" rather than "where does X happen / how does Y work", push back for a concrete question — a tour without a destination is wasted subagent load.

Sharpen the question. Rewrite it as a single sentence you can hand to each subagent verbatim. If it isn't crisp enough for a grep-guided answer, ask the user to narrow it before spawning anything.
Identify candidate clusters. Breadth first: ls -la the repo root, read the package manifest (package.json / pyproject.toml / Cargo.toml / go.mod / etc.) for entry points, then run a single keyword grep — grep -rli '<term>' <top-level dirs> — to list candidate files without opening them. Pick 2–5 plausible clusters where the answer could live. If it's obvious (single file, single function), skip delegation and read it directly.

Spawn one subagent per cluster via node "$MI_PATH" -p '<prompt>' with bg=truthy — the harness returns pid:X log:/tmp/mi-X.log and detaches the child; do NOT append &. Use this prompt template for each subagent:

Question: <the sharpened question, verbatim>
Scope: <absolute paths in this cluster only>
Do NOT read files outside the scope above.
Write your findings to /tmp/mi-explore-<cluster>.md using this exact format:
STATUS: complete | partial | blocked
SCOPE: <paths you actually read>
ANSWER: <1-3 sentences, no file:line refs here>
CITATIONS: <path/to/file.ext:<line> — <quoted excerpt>, one per line>
FOLLOW_UPS: <paths outside scope worth checking next, if any>