Sizing the Guardrail Plane So a Dead Supervisor Doesn't Leave the Swarm Unguarded
In a supervised agent swarm, guardrail capacity is the hard limit. Every action an agent proposes has to pass a policy check before it executes, so the supervisor tier has to handle the swarm's worst burst even after one of its nodes is gone. The model below takes a single calculation from the event rate to installed cores and states its error bars. It treats node loss as the event that sets the hardware budget.
The quantity that actually sets the budget
The parameter that matters is guardrail evaluation time per proposed action. That covers deserializing the intent, running the constitutional policy inside a WebAssembly sandbox, checking the swarm-wide invariants the policy reads, and returning a verdict. For the stack we assume, take it as 2 ms of CPU per event (an estimate, measured at the p99 rather than the mean, because the tail is what fills queues). The next input is the aggregate rate of proposed actions across the swarm, and that number carries most of the uncertainty. Agents don't act independently. A shared trigger, such as a tool outage or a market tick, makes hundreds of them replan at the same moment, so the rate arrives in correlated bursts and not as a smooth Poisson stream. Size against a sustained burst of 40,000 events per second (an estimate), with an error bar of plus or minus 30 percent (also an estimate), and treat the upper bound as the design point.
Carrying the numbers through
Busy cores equal event rate times evaluation time, so 40,000 events per second times 2 ms gives 80 cores fully occupied at the nominal burst (estimated). At the upper error bar that becomes 104 cores (estimated), and this is the load the surviving supervisors must absorb after a failure, not the load the full cluster shares. With four supervisor nodes, losing one leaves three, so each node needs about 35 cores of guardrail throughput below its measured queueing knee, which comes to roughly 139 cores installed (estimated). Five nodes bring each node down to about 26 cores and the total to about 130 (estimated). The redundancy overhead shrinks as nodes get smaller, but every node you add increases the cross-node state traffic needed to keep swarm-wide invariants consistent. Measure the knee on your own hardware with the real policy bundle loaded, because sandbox instantiation and policy cache misses move it more than clock speed does.
How it fails
When a supervisor node dies, its agents' events are rerouted to the survivors. If the survivors were sized against the nominal rate instead of the upper error bar, the burst pushes them past the knee and queueing delay climbs without bound. Any agent configured to proceed after a verdict timeout is then running unsupervised, and the fleet becomes least supervised exactly when it is under the most stress. Functional safety practice, in the IEC 61508 lineage, has a name for this: a protection layer whose failure mode is silent permission. It is also why the timeout path has to be designed as carefully as the evaluation path. The fix is partly hardware and partly protocol. An agent whose verdict deadline expires holds its last known safe action or degrades to a whitelisted read-only tool set. It never executes the pending proposal.
The decision rule
Install enough guardrail capacity that the nodes left after losing the largest one carry the upper-bound burst below their measured knee, make every agent fail closed when the verdict deadline expires, and accept that the remaining trade-off is between the overhead of fewer large nodes and the invariant-synchronization cost of many small ones, a cost you should measure before choosing the node count.
Technical FAQ: Why size the survivors for the upper error bar instead of adding headroom to the nominal case?
Node failures and traffic bursts are correlated. The same upstream event that makes agents replan together also tends to stress the supervisors, through retries, cache churn and reconnect storms, so the failure case and the peak burst usually arrive at the same time and have to be sized together.