Agentic search burns 30-50% of agent tokens: RL-trained model is 20x faster and 100x cheaper
AI Engineer · youtube · 2026-09-17
On the AI Engineer channel, Maximilian-David Rumpf of SID.ai explains why search is ripe for RL.
The problem
- Some 30-50% of an agent's tokens go to searching, almost all up front.
- Handing a task to an agent roughly doubles the odds of finding the right documents vs a search engine — at 100-1000x the cost and minutes instead of milliseconds.
Why classical pipelines can't close the gap
The rewrite→retrieve→rerank→return pipeline freezes every decision at design time with a fixed compute budget per query; the reranker often knows results don't answer the question but can't act, producing a long tail of failures patched with endless edge cases.
Why RL fits
- Verifiable reward: you either found the right document or not.
- Grindable environment: thousands of attempts per second.
- Follows the path of computer vision and chess: learned design beats human design.
Results
A specialized model runs 20x faster than a frontier model on the same task (5s vs 2min) at 1/100th the cost, keeping bad results out of the main context.
More from coding & agent
- Microsoft's MemSteer turns unlabeled user histories into agent memory supervision — MicrosoftResearch · 2026-11-02
- She fired herself from her own Instagram: AI agent runs the account, views up 82.3% — heyshrutimishra · 2026-11-02
- Sourcegraph: AI writes code faster than teams can review it, driving more outages — AI Engineer · 2026-11-02
- Tsinghua & Microsoft's HarnessMM cuts agent token use by up to 87.93% while raising success rate — rohanpaul_ai · 2026-11-02
- GitHub launches Agent HQ: Anthropic, OpenAI, Google coding agents in Copilot — emmanuelvivier · 2026-11-02
- DeepSeek open-sources Super-Agent, a coding agent that reads team chats to minimize human input — emmanuelvivier · 2026-11-02