Parallel agents are useful when the work can truly be decomposed. They cost more tokens and coordination, so use them for independent exploration, testing, triage or evidence gathering—not for several agents editing the same thing at once.
1. Break the task into independent questions with non-overlapping deliverables
Split the research by questions that can be answered independently, not by arbitrary page counts. Good partitions produce different evidence sets, such as regulation, market data and implementation constraints.
2. Assign one role and evidence standard to each subagent
Give every subagent the same evidence contract: preferred source types, date range, geography, citation requirements and how to express uncertainty. This makes later comparison meaningful.
3. Keep irreversible actions and shared-file edits in the main thread
Keep shared-file edits and consequential external actions in the parent task. Subagents should primarily inspect, calculate or draft isolated findings so parallel work cannot overwrite one another.
4. Wait for all requested subagents, then compare overlaps and disagreements
Wait for all branches, then compare where they agree, disagree or use incompatible definitions. A disagreement is a signal to inspect source quality or assumptions, not something to average away.
5. Ask the main thread to synthesize only after conflicts are resolved or clearly labeled
Synthesize only after conflicts are resolved or explicitly labeled. The final report should preserve provenance so a reader can trace a major conclusion back to the branch and source that produced it.
Use parallel subagents for this research. Agent A: [question]. Agent B: [question]. Agent C: [question]. Each agent must return evidence, source dates and uncertainties, and must not edit shared files. Wait for all agents, compare disagreements, then produce one synthesis with a conflict table.
Parallelizing dependent steps; giving every agent the same vague prompt; allowing several agents to write the same file; merging findings before checking contradictions; using subagents for tiny tasks that are faster serially.
Finish check
Each agent has a distinct scope, the main thread can trace every important claim to evidence, and unresolved disagreements are visible instead of being silently averaged away.