Multi-Agent Analysis

Steps

Step 1 — Frame the problem

• Frame: restate the problem in a line or two and name its type — verify a claim/fix · which answer is correct · architecturally correct + simplest · how something works.
• Task list: if the ask has independent parts, list them first (TaskCreate) so none is dropped.

Step 2 — Size and dispatch

• Size: Use 1 subagent by default for a focused question or when scope and complexity are limited. Use 2–3 subagents only when the problem's scope, size, or complexity is large enough to create independent angles worth analyzing in parallel.
• Dispatch: send all selected Agent calls in one message. Model per task — opus for architectural/design judgment, sonnet for mapping, code-tracing, or doc/web research.
• Brief each:
  • What to pass: the problem statement, its own angle, and file paths (not pasted contents).
  • Ground every claim in source it read.
  • Return the Findings schema, not a raw dump.
  • Flag anything material it notices outside the question.
  • Don't fan out — no subagents of its own; keep the tree one level deep.