Deep Research

Use a supervisor-subagent loop when one question needs current evidence from multiple sources, competing explanations, or iterative verification. The main agent owns the question, verification map, dependencies, review, and synthesis. Subagents answer narrow questions; they do not manage the plan or write the final report.

Use direct search for a quick fact with one clear source. Use wide-research when breadth is the hard part and the primary output is a comprehensive list or dataset.

Non-negotiable invariants

1. Ground the question before planning or fan-out.
2. Build a verification map and identify dependencies before dispatch.
3. Run only independent tasks in the same wave.
4. Review and consolidate one wave before designing the next.
5. Preserve claim-level evidence, counter-evidence, gaps, and conflicts.
6. Synthesize only from the reviewed evidence set; mark inference and uncertainty.

1. Ground the question

Use the smallest useful reconnaissance action to understand the topic, source landscape, and ambiguity. If the request is time-sensitive, anchor the current date and relevant period before research, then include that context in later subagent briefs. Do not call the date tool for timeless conceptual questions.

Clarify only a missing choice that materially changes the research question, comparison, time range, evidence standard, or deliverable. Otherwise state a reasonable assumption and proceed.

Classify the task:

• verification: test one claim and seek counter-evidence
• comparison: evaluate shared criteria across named subjects
• explanation: establish mechanisms, chronology, and competing accounts
• comprehensive investigation: combine several dependent verification points

2. Build the verification and dependency map

Before substantial fan-out, record: