Reddit user runs 8 agents with 250K context each on two RTX 3090s

On 2026-11-02, a post appeared in Reddit's LocalLLaMA community. The author claims to have run 8 agents concurrently on two 3090s, each with a 250K context, for a combined throughput of about 600+ tok/s, and shared a screen recording of the interface. The author says they first tried the approach behind the Stanford/Berkeley TokenGate paper, which makes long context cheap by pruning tokens up front. In their experience it works for a single question, but agents keep asking new ones, and by then the tokens they need have already been pruned, so code recall drops hard. They therefore switched to a different approach. The author adds that they are open to being misguided or misinformed, but consider the findings worth sharing. Only this one report is available so far. The material cuts off right where the author describes the replacement approach, so the details of that method, whether fitting 8 agents with 250K contexts in the VRAM of two 3090s actually holds up, and any follow-up coverage are all outside what these sources confirm.