Classic retrieval systems are hand-engineered: a human decides when to search, how to phrase the query, and when enough is enough. The 2026 shift is to stop hand-writing that logic and instead train the agent to discover good retrieval behavior on its own.
From hand-crafted to learned
The idea is to treat retrieval as a policy the agent learns through trial and reward. Given a goal, the agent decides when to search, what query to issue, whether the results are sufficient, and when to act versus search again — and it's rewarded for reaching correct outcomes. Over training, it discovers retrieval strategies a human might never hand-code.
Don't tell the agent how to search. Reward it for finding the answer, and let it work out the how.
Why it beats hard-coded pipelines
A fixed pipeline searches the same way every time, whether or not that fits the question. A learned policy adapts: it searches more aggressively on hard questions, rewrites a bad query, and stops when it has enough. That flexibility is exactly what static retrieval lacks — and it's why retrieval is becoming a loop the agent drives rather than a step it passes through once.
The bigger picture
This is part of a broader pattern: replacing hand-tuned agent scaffolding with learned behavior wherever a clear reward exists. Retrieval has an unusually clean signal — did the retrieved information lead to a correct answer? — which makes it a natural place to apply it. The result is agents that treat search as a skill to be exercised judiciously, not a reflex to be fired on every step.
