Fan out sub-agents in parallel; use Haiku for the noisy ones

When the current task involves any of: scanning many files, web research across multiple URLs, scraping/reading >50K tokens of source material, exploring an unfamiliar codebase area, or comparing N options, don't read it all into the main thread. Spawn parallel sub-agents via the Agent tool. Use Haiku for the cheap noisy ones, reserve Opus/Sonnet for synthesis.

Why: the main thread is precious context. Pulling 100K tokens of articles, source files, or web fetches into Opus to summarize them once is wasteful and burns subscription quota faster than necessary. Haiku is included in the Claude Code subscription quota and reads big inputs cheaply. Sub-agents return only their summary to main, keeping the conversation lean.

How to apply: - Send multiple Agent tool calls in the same message when the work is independent (the docs explicitly say this is correct usage). - Pick Explore (Haiku-grade) for known-target code lookups; general-purpose (Sonnet) for ambiguous research; worker for one-shot fixes. - For Telegram bot brains and dispatcher pipe-mode jobs running claude -p, default the model to Haiku unless the prompt explicitly needs Sonnet/Opus reasoning. Long article triage = Haiku. Code planning = Sonnet. System architecture = Opus. - Don't fan out for trivial tasks. The bar is "would I read >20K tokens just to extract a summary?"

Validated by: Hack #11 + #13 in Nate Herk's "32 Tricks to Level Up Claude Code" (memory/daily/2026-05-01.md checkpoint), aligned with feedback_subscription_over_api_for_automation (Haiku is subscription-billed, never Anthropic API).