Skip to content
FlintFlintby Torch
← LevelsBuilder · B9

Summaries, not raw text

Build-along · ~8 min at your own keyboard

The classic fan-out failure: ten agents each return their full working text, the parent's context explodes, 'prompt too long.' Rule: subagents return conclusions — counts, exceptions, a verdict — never their raw material.

The long version: Subagents: return summaries ↗

Do this

0/3

Checkpoints

  • ▸You should now see a written return contract with exactly what each agent sends back.
  • ▸You should now see short returns: a count, exceptions and a status line.
  • ▸You should now see a parent context that stayed small after the run.
The three ways this goes wrong
  • The agents send back long text anyway. — Put the contract at the end of each agent's prompt and say 'nothing else'.
  • The parent shows 'prompt too long'. — Cut the chunk size or the number of agents, and make agents write detail to a file instead of returning it.
  • You cannot tell whether an agent finished. — Add a required DONE or FAILED word to the contract.

Share what happened

Paste your return contract and what the agents actually sent back.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.