Results · the fan-out team

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

Every agent is Qwen3-4B, on one H100 GPU. The test is FanOutQA: 310 questions that fan out: each needs facts from many Wikipedia pages, such as "Which companies are worth over a trillion dollars, and when was each founded?"

Three systems

One agent against two orchestrator-workers teams

One agentone model reads every page itself, one after another

Team in wordsthe orchestrator writes the plan; each worker thinks in written words, then writes its report to the reader

Latent teamthe same team; each worker thinks in vectors instead of words and hands the reader those vectors and a three-line written report (the entity, the fact, the page title); the plan and the answer stay written

Role	Team in words	Latent team
Orchestrator	thinks in words, writes the plan: the list of pages to look up	the same: thinks in words, writes the plan
Worker	reads its page, thinks in words, writes a report to the reader	reads its page, thinks in vectors (2 to 8 of them, it picks when to stop), hands the reader those vectors and a written report of at most three lines: the entity, the fact as the page states it, the page title
Reader	reads the reports, thinks in words, writes the answer	takes each worker's vectors in as if they were tokens, then its three lines; thinks in words, writes the answer

One question

The latent team answers first and spends the least compute

Played at 2x 0.0s

Metrics

The latent team against the team in words

The result in one picture

Up is more accurate, left is faster, smaller spends less compute

Bigger teams

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

One agent against two orchestrator-workers teams

The latent team answers first and spends the least compute

The latent team against the team in words

Up is more accurate, left is faster, smaller spends less compute

More lookups need more agents; the latent team stays fast and cheap as the team grows