Research · 2026-07-08
Measuring a fleet: agentctl control-plane benchmarks
Published numbers for the questions that matter operationally: per-agent overhead, work-distribution throughput, exactly-once claims under contention, scale-from-zero latency, and reconcile behavior at fleet size.
Infrastructure claims are cheap; agentctl publishes its receipts. All numbers below were measured on a single-node kind cluster on an AMD EPYC 7502P, with methodology and raw runs in the repository.
Idle cost. An idle managed agent costs ~1.3 millicores and under 1 MiB of control-plane memory. The control plane itself idles around 16 millicores and 76 MiB across 7 pods — small enough to run on the cluster you already have.
Work distribution. The claim-based work API sustains ~5,100 operations per second with p99 under 100 ms — measured at zero errors across 41,000 operations. Under 72 concurrent claimants hammering the same queue, the double-grant count is exactly zero: leases are exactly-once by construction, not by luck.
Elasticity. Scaling a claim fleet from zero to five running agents takes about 2.2 seconds end-to-end — fast enough that scale-to-zero is a real default, not a cost-saving compromise. Node packing tops out around 82 agents per node, kubelet-capped rather than agentctl-capped.
Flat reconcile. Reconcile latency stays flat from 1 to 100 managed agents — the controller's work per agent does not grow with fleet size, which is the property that separates a control plane from a script.
The benchmark page exists for a second reason beyond marketing honesty: it pins a regression baseline. Every release re-runs the suite, and a significant delta is release-blocking. Fleet software earns trust by publishing the same numbers twice.
Canonical: agentctl repo · docs/benchmarks.md (methodology + raw runs)