Invest1 publisher3 min readPublished
Fed builds a rule book for the AI agents that now make half of FRED's visits
Fed Governor Christopher Waller said AI agents now account for about half of visits to FRED, a site whose traffic is growing 150% a year. The Fed is adapting its flagship database for the software that now sits between official data and the people who use it.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Waller said the Fed will soon roll out a model context protocol, a digital rule book that lets AI agents and outside tools communicate with its databases.
- The Fed is also working to open its FRASER library and its ALFRED archival database to AI inquiries.
- So far, according to Waller, the bots reading FRED have generally been very accurate, though problems may arise without an interoperability interface.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- exposure Half of FRED's audience now reads it through software, so an agent's unstated assumption can reach its user attached to an official series.
- capability One documented interface covering more than 850,000 series, other agencies' data included, would give agent builders an approved way to pull and cite official numbers.
- decision Firms building research agents will have to choose between the Fed's protocol and pages designed for humans, which Waller says bots already handle well.
Read together, Waller's two traffic figures say something about the human visitors as well. Growth of 150% a year means FRED now gets 2.5 times the visits it had a year earlier [2]. People make up roughly half of today's total [1]. So human visits alone equal 1.25 times the whole site's traffic a year ago, which means they rose at least 25% even if every earlier visit came from a person [1]. That takes the 150% as the past year's rate.
According to Waller, most of the growth comes from machines [2]. He was director of research at the St. Louis Fed before joining the Board in 2020 [14], and he gave the speech there [3]. "Up till now, FRED has been designed to be viewed and understood by humans. But an AI bot sees and interprets data very differently," Waller said. "And that means the FRED team must adapt how FRED is presented to this new type of viewer so that the data are conveyed back to the agent's master accurately and in a well-documented manner." [4] At 150% a year for two more years, the site would carry 6.25 times today's traffic [2]. The work Waller described is all on the machine side: the protocol for FRED, plus opening the FRASER library and the ALFRED archival database to AI inquiries [9].
Waller called FRED "the greatest public good that we have ever created" [13]. An official series is worth something to an investor because it is the official number, defined the way the agency defines it. If an agent paraphrases that series with an assumption it does not state, its user gets a different number under the Fed's name. Waller said the protocol is meant to get FRED data cited accurately and reproduced with as few hallucinations, biases and hidden assumptions as possible [6].
FRED has changed format before. Over 60 years it went from a weekly newsletter to a telephone answering machine, then a computer bulletin board, then a website [12]. Waller said this change is different. "Unlike past technological advances that primarily aided in the distribution of data, AI can create content as well, and this capability creates new risks," he said [7]. He also said the bots have so far "generally been very accurate in describing visualizations and explaining complex terms and concepts in simple ways" [8]. The protocol targets problems that he said may arise without an interoperability interface [8].
I see three ways this can go. Agent builders could route their queries through the protocol, and answers about US data would converge on the Fed's documented version. Builders could also keep reading the pages built for people, which Waller says the bots already describe well [8], and leave the protocol lightly used. Or FRED gets its interface while the archival tools lag behind. I think the first is likelier. One interface would reach more than 850,000 series, including data FRED gathers from other government agencies and global sources [10], and a documented route is cheaper for a builder to maintain than a homemade one. The case against is the accuracy Waller already credits the bots with: an agent that already works has little reason to change how it reads the site.
Waller described FRED's users as analysts and decisionmakers in the public and private sectors [11]. The reported remarks do not say who runs the agents, so nobody outside the Fed knows how much of the bot half works for investors. The view is wrong if the bot share keeps climbing past half while the protocol carries only a small slice of that traffic.
What to watch
- A launch date for the FRED protocol and the list of series it covers at launch, which Waller described only as coming in the near term.
- Any reported case of an AI agent misquoting a FRED series, the risk Waller named as the reason for the protocol.
- Whether the government agencies whose data FRED aggregates build agent interfaces of their own.