Literature database · Cost-efficient multi-agent LLM systems

Cost-Efficient Agents Literature

A searchable index of … papers, preprints, docs, blogs and repos on cutting the cost of production multi-agent LLM systems without losing quality: query- and step-level routing, cascades, config optimization, learning policies from logs and live traffic, quality signals, distillation, caching and serving. Snapshot of 2026-09-30.

Try:
How to read relevance, granularity and cost numbers
  • Relevance 1–5 rates how central an item is to the premise, not paper quality. 5 = cost-aware routing or config optimization in agentic or multi-agent systems; 4 = query-level routing or cascades, clearly transferable; 3 = supporting technique (caching, compression, distillation, judges); 2 = background. Scores come from the curating agent's reading; treat one-point differences as noise.
  • Routing granularity says where the decision is made: query (once per request), step (per turn or action inside a trajectory), subagent (per role or agent), token (inside decoding), or none for non-routing work.
  • Cost numbers are only comparable on the same benchmark, baseline and price sheet. "2x cheaper at 95% of GPT-4 quality on MT-Bench" and "72% lower cost on SWE-bench" are different quantities. Numbers are as the authors state them; nothing was re-measured. Vendor posts are marked vendor-reported.
  • Production evidence is set only when real deployment or real traffic data is described.