Results

Two measurements of agents that hand over the model's memory instead of words, against the same agents writing to each other. Every agent is Qwen3-4B on one H100 GPU; every number is the mean over three to five seeded runs.

1 · A team on fan-out questions

Workers that think in vectors make a team faster and cheaper, as accurate

One question needs facts from many Wikipedia pages at once. An orchestrator splits it, one worker a page looks a fact up, a reader writes the answer. The same team twice: workers that think and report in words, and workers that think in vectors and hand the reader those vectors with three written lines. One agent reading every page itself is the reference.

Seconds a question on 310 questions, the share of facts right at the end of each bar.

2 · A chain of eight readers

At the same size, a chain that hands over its memory keeps twice what written notes keep

Eight copies of the model stand in a line. Each reads one Wikipedia paragraph and hands on to the next; the last answers questions on all eight, and no agent in the middle knows the questions. The handoff is a written note or the sender's working memory cut to the note's size, at six sizes from 64 to 2,048 tokens or positions. The whole memory and one agent reading all eight paragraphs are the ceiling.

Score after seven handoffs by the handoff's size, tokens for the note and positions for the cut memory; two rules cut it, neither sees the questions.