More compute made one agent 13x faster, another just 4%: agents' bottlenecks are task-dependent
alex_verem · x · 2026-09-24
A Korea University and Microsoft Research Asia paper measured where agents spend time across RAG QA, web search, and coding tasks. Key findings:
- Web search agents use only 14% CPU, waiting nearly half their time on websites; QA agents burn 62% of time searching local libraries at 86% CPU
- Going from 2 to 8 cores made the QA agent 13x faster but the web search agent just 4% faster
- Tool concurrency limits sped up the QA agent 3.2x while slowing the coding agent 2.8x
- Redistributing the same 6 cores cut average response time 32% with no new hardware (QA agent 5.4x faster)
Takeaway: before buying more compute, check what your agent is actually waiting on.
More from coding & agent
- Stealth Model Space Bunny Free on OpenCode: One-Prompt Full Game, 1M Context, Zero Retention — iamfakhrealam · 2026-09-24
- Depthfirst Uses Agents to Crack Dependency Reachability, Beating Traditional Systems — andreamichi · 2026-09-24
- Dev launches free no-setup GSC MCP server after finding Google data can be wildly misleading — StatisticianFormal44 · 2026-09-24
- GPT-6 coding agent regression: clones entire repos instead of switching branches — banteg · 2026-09-24
- Together launches Jev-style Tev1-4B classifier on Qwen3.5, train your own for $17 — usamawahabkhan · 2026-09-24
- Open-source tool bough visualizes Claude Code tasks with per-task token cost from local history — VoidEqualZero · 2026-09-24