Back to Details

parallel-agents

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Parallel Agents

Parallel Agents

Steps

步骤

  1. Slice lanes. Read-only first: one agent per independent question, directory, package, axis. Launch concurrently.
  2. Prompt tightly. Give each agent: lane, read/write mode, allowed paths, forbidden paths, expected output shape.
  3. Write only when non-overlapping. Fixed-contract, non-overlapping file sets. Two lanes might touch same file: keep sequential.
  4. Keep moving. While agents run: do local reading, planning, small safe edits. Do not poll.
  5. Integrate. Read summaries. For writers: verify diffs. Resolve seams in main thread. Run narrowest useful check.
  1. 拆分通道：先处理只读任务：为每个独立问题、目录、包、维度分配一个Agent，并发启动。
  2. 精准提示：为每个Agent提供：通道信息、读写模式、允许路径、禁止路径、预期输出格式。
  3. 仅在无重叠时写入：采用固定约定的无重叠文件集。若两个通道可能涉及同一文件，则保持顺序执行。
  4. 持续推进：在Agent运行期间：执行本地读取、规划、小型安全编辑操作。不要轮询。
  5. 整合结果：读取摘要。对于写入型Agent：验证差异。在主线程中解决衔接问题。运行最必要的检查。

Rules

规则

No duplicate lanes. No agents for one-line lookups. Never delegate ownership. Prefer more read agents over more write agents.
禁止重复通道。不要为单行查询创建Agent。绝不委托所有权。优先使用更多只读Agent而非更多写入Agent。