Rule of Thumb: Each Sub-agent Costs 2x Tokens for ~50% Latency Gain
brandon_galang · x · 2026-10-09
An Anthropic engineer shares a practical rule of thumb for sub-agents: each extra agent roughly buys 50% latency speedup at 100% extra token cost. Smaller models can shift the cost math but think and work slower — sometimes plain fast mode is cheaper. Parallelism also has ergonomic benefits beyond raw speed.
More from coding & agent
- Manning publishes 'Building LLM Applications with DSPy' on systematic LLM development — lateinteraction · 2026-10-10
- StepFun's Step 5 Preview free in Kilo for a week: 600B params, 27B active, 1M-token context — StepFun_ai · 2026-10-10
- YC partner shares multi-agent workflow: agents write briefs for each other like middle managers — ycombinator · 2026-10-10
- AI writing skill boosts output to 4 posts a week, human editors become the bottleneck — danshipper · 2026-10-10
- AI agent digs through Azure billing to recover nearly $2,000 in lost credits — pswider · 2026-10-10
- HQ Agent Tool Ships Major Update: Per-App Postgres, Full-Stack Next.js, Interactive Pages — jacob_posel · 2026-10-10