Research
We write up what we learn while building our products. Every paper is free to read on arXiv.
ObjectGraph (paper on arXiv, opens in a new tab)
A file format that lets AI agents look up the part of a document they need instead of reading the whole thing. In our tests it used up to 95.3% fewer tokens without hurting accuracy.
Phase-Scheduled Multi-Agent Systems (paper on arXiv, opens in a new tab)
Agents take turns instead of all running at once, and the ones waiting get a short summary rather than the full context. That used 27.3% fewer tokens, with results within 2.1 points of running every agent.