2026-07-2809:151 min readby Shai Chikorel
Subagents, properly: when forking context beats one long thread
A practical model for when to spawn a subagent, what to put in its prompt, and how to keep the parent thread cheap.
#agents#workflow#context
// key takeaways
- ›Spawn a subagent when the work is read-heavy and the findings compress well.
- ›Keep the parent thread as the plan; let children do the searching.
- ›Return a short structured report, not a transcript.
Why fork at all
A single long thread pays for every file it ever read. A subagent reads the same files in its own context, then hands back a summary a fraction of the size.
The rule of thumb
Fork when input tokens are large and the useful output is small. Exploring a codebase, auditing dependencies, scanning logs — all ideal.
# ask for a report, not a narration
claude -p "Find every place we read process.env directly. Return a table: file, line, var."
What to put in the subagent prompt
- The exact question, with the shape of the answer you want back.
- The boundaries: which folders, which file types.
- A hard instruction to return findings only.
Keeping the parent cheap
Treat the parent thread as a plan document. Paste the report, act on it, and
/clear once the task is done — it is cheaper than /compact when you no
longer need the history.