What is backpressure in Agent?
What should be done?
On the surface, the system keeps busy; fewer and fewer successful results are actually accomplished.
From a runtime perspective, backpressure reflects that feedback is not transmitted back to the upstream in time: the downstream is saturated, and the scheduler continues to dispatch tasks.
When troubleshooting this type of problem, look at two points: How many successful results are really completed per second?
How many requests does it take to get a successful result?
The control method is actually very simple: • Use Semaphore to limit tasks in transit at the same time • Use bounded queue to prevent infinite accumulation of tasks • Use exponential backoff + jitter to break up the retry peak • Scan the goodput curve under different concurrencies to find the sweet spot Multi-Agent scheduling is becoming more and more like a small distributed system.
The real trouble is that this sweet spot keeps moving with model latency, token quotas, network fluctuations, and downstream load.
This video explains it very clearly and deserves to be repeated three times!
When Agents are deployed concurrently to 100, why are only 32 successful?