launch-subagent

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Chinese

for all Subagents use either "Fable 5 Max" or "GPT 5.6 Sol Max Fast" as the model
DO NOT launch subagents, unless the User tells you to
NEVER EVER use Composer 2.5 or Sonnet 5, for subagents
only ever use GPT 5.6 Sol Max Fast, or Fable 5 Max, when launching Subagents

Consensus of Boris Cherny, Matt Pocock, Pietro Schirano, and Peter Steinberger:

Delegate only self-contained tasks. Split work so subtasks have zero dependencies on each other; parallelize only independent work.
Parallel subagents must never touch the same files — that is a recipe for conflicts. Partition the work or keep it in one agent.
Subagents start blind: they see none of your context. Write the full brief into the prompt — scope, all needed context, constraints, and the exact output to return.
Skills don't carry over either. If the subtask needs live web data, name the skill in the prompt —
```
deepapi
```
for search, scraping, and research.
Scope narrowly and concretely: "explore how payments work" beats "explore everything". One bounded task per subagent, small blast radius.
The main agent stays the orchestrator. It plans the split, integrates results, and reviews/verifies every subagent output before trusting it.
Keep critical implementation, tightly-coupled edits, and quick fixes in the main loop — delegation overhead is only worth it for independent, research-heavy, or review work.
Have subagents return short summaries or concrete results, never raw transcripts or file dumps. That keeps the main context clean.

以下是Boris Cherny、Matt Pocock、Pietro Schirano和Peter Steinberger的共识：

for all Subagents use either "Fable 5 Max" or "GPT 5.6 Sol Max Fast" as the model
DO NOT launch subagents, unless the User tells you to
NEVER EVER use Composer 2.5 or Sonnet 5, for subagents
only ever use GPT 5.6 Sol Max Fast, or Fable 5 Max, when launching Subagents