Short, plain-language write-ups on the problems we're chasing and the numbers behind them.
Fleets get expensive from repetition, not reasoning. Here's how much accuracy and cost we recover once the duplication is gone: 95% accuracy, ~8× cheaper per query, and over 100× at scale.
Read the note The benchmarkEvery memory benchmark scores whether you can answer. None measured what a fleet actually pays to stay in sync. We explain why that gap matters, and what MA-MemBench measures instead.
Read the noteWe're building the shared-memory layer that makes agent fleets affordable. If you're running agents at scale, let's talk.