Sam Green, team lead for applied AI at Bridgewater Associates, told AI Engineer that the firm's Vault turns more than 100 million documents into an investing edge and showed how a generic web agent failed his test on AI infrastructure and the Korean won.
Aaron Linsky, who runs product engineering at Bridgewater, said the hard problem was building one platform for two different readers.
He said humans read at about 300 words per minute and need high precision, while agents read at 30,000 words per minute and will attend to anything you hand them, so the team benchmarked precision at 10 and found Vault at about 60 percent before fixes.
Green showed the web agent found 15 sources but cited one or two and leaned on a July note to answer an early October question, while the Vault-backed agent scoped Oct. 1 to Oct. 8, surfaced 52 candidates and cited seven with enrichments for country, concept, market and time period discussed.
Linsky said classic post-filtering was fast but lost recall and pre-filtering kept recall but was slow, so Bridgewater switched to Turbo Puffer native filtering that scores only nearest clusters with matches and paired it with infra that handles about 100 queries an hour for 300 humans versus 100 plus queries per second for agent swarms.
He argued quality alone did not change habits, so the team rebuilt email workflows as saved searches and saved collections inside Vault, then put Vault in chat, Slack, a phone app, the IDE and a CLI for agents to power a pocket analyst that does hours of work in minutes and a forecaster that turns unstructured text into macro predictions.
Green said the prize now is an investor workbench where agents run proactively, surface cards with summaries and evidence, and write the collaboration back into Vault to compound the next turn, but neither he nor Linsky explained how entitlements, governance or point in time safety hold up as thousands of documents arrive minute by minute.
