Disable subagents to save tokens: field guide to single- vs multi-agent cost setups
dumpshoot · reddit · 2026-10-12
A practical breakdown of usage-limit-friendly agent setups: for Codex, stick to one agent and disable subagents — agents don't trust each other yet and burn tokens double-checking each other's work. For Claude, options include a single agent, on-demand advisor mode, or smart-model orchestrator + Haiku subagents; watch for Haiku's 2x cost past 100k context and the 5-minute default subagent cache (1-hour cache costs 60% more to write). The author tests Opus 5.5 as orchestrator picking between cached Haiku variants, finishing two hours of work using only 5% of the $20 weekly allowance. Post includes a full single-file WebGPU "Plush Hammerhead" prompt with XPBD soft-body simulation.
More from coding & agent
- MetaKernelBench measures cross-DSL GPU kernel skill transfer, with lifts from -19% to +29% — encodelab · 2026-10-12
- Developer's Grok trading bot keeps running but hasn't made much money yet — nima_owji · 2026-10-12
- ACT-R memory equation vs. simple recency decay: crossover comes at 2.7 hours, not much gain — memokris · 2026-10-12
- Anomalous Model Browser: free ComfyUI plugin matches models by hash and relinks missing files — demongatanjieu · 2026-10-12
- ChatGPT Work agent drafts and sends emails in 76 seconds while owner is fishing — gabrielchua · 2026-10-12
- Indie dev's multi-agent setup lets Codex give Claude Code instructions directly — AlchainHust · 2026-10-12