念
Parallelization
Parallelization — multiple LLMs work on the task concurrently; outputs are merged programmatically
Two flavors of parallelization: Sectioning splits a task into independent subtasks running concurrently; Voting runs the same task multiple times to get diverse outputs and take consensus. Shared property: subtasks are known at design time, not decided at runtime (distinguishing it from Orchestrator-Workers).
Sectioningfits: task splits cleanly + throughput is the bottleneck
user-query handling + content guardrail review (concurrent)multi-file codebase analysis in parallel
Votingfits: higher confidence needed + multi-angle review
multiple prompts review code for security holes from different anglesmulti-sample consensus (Self-Consistency)
Place in agentic systems vs Orchestrator-Workers parallelization subtasks are predefined; Orchestrator-Workers decides them at runtime Cost–benefit low overhead (no extra orchestration call); suits clean-structure batch work with independent subtasks
→ Orchestrator-Workers · Agentic Systems · GuardrailsAnthropic (2024)
Parallelization（并行化）
定义
多个 LLM 同时处理任务，结果由程序汇聚。有两种变体：
- Sectioning（分段）：把任务拆成独立子任务并行执行
- Voting（投票）：同一任务跑多次获得多样化输出，取共识
适用场景
子任务可独立执行时用 sectioning 加速；需要更高置信度或多角度审视时用 voting。
典型用例：
- Sectioning：一个模型处理用户查询，另一个同时做内容审核（guardrail）
- Voting：多个 prompt 从不同角度审查代码安全漏洞
在 agentic 系统中的位置
属于 agentic systems 中的 workflow 模式。与 orchestrator-workers 的区别：parallelization 的子任务是预定义的，orchestrator-workers 的子任务是动态决定的。
References
sources/anthropic_official/building-effective-agents.md