AI Agents for Economic Research: 6 Jobs for Economists and Macro Teams

Economists spend much of their time getting numbers: the latest CPI print and its components, cross-country forecasts, long series for a chart, what markets expect for the next release, and the newest papers on a method. That is where AI agents for economic research help most. Anton Korinek's NBER working paper "AI Agents for Economic Research" shows economists how to build them, starting with an agent that fetches data from FRED. In practice, AI agents for economic research work when they can reach the official series, forecasts, prediction markets and papers directly, and return every number with its series ID and date.

This guide follows six jobs, each for a different economist or macro analyst.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for economic research do

An AI agent for economic research is a language model that can search official statistics, forecasts, prediction markets and research papers, read what it finds and return cited, structured answers. It doesn't set policy or forecast for you. It does the data gathering and first reading that come before an economist's judgement.

Figure 1. Six economic research jobs we ran on November 9, 2026, with the sources used, what came back and what to do next.

1. Release-day brief: the numbers before the print

The problem: a sell-side US economist is writing tomorrow's CPI preview and needs the latest headline and core inflation, from the official series rather than news coverage.

What we ran: two searches of the BLS dataset, one for headline CPI and one for core.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
headline = valyu.search(
    "US consumer price index all items 12-month percent change September 2026",
    included_sources=["valyu/valyu-bls"],
    max_num_results=3,
)
core = valyu.search(
    "CPI all items less food and energy",
    included_sources=["valyu/valyu-bls"],
    max_num_results=3,
)
series = headline.results[0]
print(series.metadata["series_info"]["series_id"])  # which BLS series came back

What came back: two monthly series with their BLS IDs: the seasonally adjusted all-items CPI (CUSR0000SA0) from January 2025 to September 2026, and seasonally adjusted core CPI (CUSR0000SA0L1E). From these, headline CPI rose 3.7% in the year to September 2026 and core 2.6%.

What we checked at the source: the BLS API. The unadjusted all-items index (CUUR0000SA0), which BLS uses for the headline 12-month change, rose 3.74% in the year to September, and unadjusted core 2.56%. Both round to the same figures as the adjusted series.

The catch: two of them. First, the search returned the seasonally adjusted series; for 12-month changes, quote the unadjusted series, as BLS does, and say which one you used. Second, October 2025 has no value. BLS marks it "Data unavailable due to the 2025 lapse in appropriations", and Valyu returns it as null. An agent that computes tomorrow's 12-month change from the index will hit that gap.

The decision: use the figures and notes in tomorrow's BLS release for October's 12-month change, and quote September's 3.7% headline and 2.6% core from the unadjusted series in the preview.

2. Market expectations: what traders price for the next print

The problem: a macro trader wants to know what prediction markets expect for October's CPI before the release at 8:30 a.m. Eastern on November 10.

What we ran: a search of Valyu's Kalshi dataset.

markets = valyu.search(
    "October 2026 CPI inflation",
    included_sources=["valyu/valyu-kalshi"],
    max_num_results=5,
)
for market in markets.results:
    print(market.title, market.url)

What came back: five Kalshi CPI events with implied odds, volume and rules. Four settle on the October release, closing around 13:29 UTC on November 10; the fifth asks how high CPI will get in 2026. The October market with the most volume priced the month-on-month change like this:

What we checked at the source: the rules and resolution times in each returned record, which link to the Kalshi event pages.

The catch: each market settles on a different number. The main market resolves on the monthly change to one decimal place; another resolves on the unadjusted CPI-U index level; none uses the 12-month change in job 1. Read each market's rules before quoting it, and record the time you read the odds, because they move until the release.

The decision: frame the preview around a 0.2% to 0.3% monthly rise, with the market centred close to the 0.2% threshold.

3. Cross-country outlook: forecasts and actuals side by side

The problem: a sovereign analyst comparing Germany, France, Italy and the UK needs last year's growth and the IMF's projections for 2026 and 2027.

What we ran: a search of IMF data and a search of World Bank indicators.

forecasts = valyu.search(
    "real GDP growth projections Germany France Italy United Kingdom 2026 2027",
    included_sources=["valyu/valyu-imf"],
)
actuals = valyu.search(
    "GDP growth annual percent Germany France Italy United Kingdom",
    included_sources=["valyu/valyu-worldbank-indicators"],
)

What came back: IMF real GDP growth (NGDP_RPCH) and World Bank GDP growth (NY.GDP.MKTP.KD.ZG) for all four countries:

What we checked at the source: the IMF DataMapper API. The IMF figures match.

The catch: the two sources disagree on the UK's 2025 growth, 1.4% against 1.2%. Different publishers update on different schedules and revise at different times. Take actuals and forecasts from one source where you can, and note which edition you used.

The decision: use the IMF series for both 2025 and the projections, so the comparison is internally consistent, and footnote the World Bank figure for the UK.

4. Long series for a chart: ask for the frequency you need

The problem: a research assistant needs the 10-year Treasury yield since 2000 for a chart in a policy paper.

What we ran: a search of FRED that named the frequency and the start year.

response = valyu.search(
    "10-year Treasury constant maturity rate monthly since 2000",
    included_sources=["valyu/valyu-fred"],
    max_num_results=1,
)
info = response.results[0].metadata["series_info"]
print(info["series_id"], info["frequency"], info["count"])

What came back: FRED series GS10, the monthly 10-year yield, with 322 observations from January 2000 to October 2026, about 33,000 characters. October 2026 averaged 5.13%. The record also carries the series' last update, November 2, 2026, and its vintage date.

What we checked at the source: FRED. GS10 for October 2026 is 5.13%, and the recent months match.

The catch: in an earlier search for the "US 10-year Treasury constant maturity yield", with no frequency or start date, FRED returned the daily series DGS10: 16,918 observations back to 1962, about 1.7 million characters. That is far more than a chart needs, and far more than a model's context should hold. Name the frequency and the start date in the question.

The decision: chart GS10 from 2000, cite the series ID and the date it was retrieved, and switch to DGS10 only for daily moves around specific events.

5. Recession risk: read the odds with the volume

The problem: a strategist wants a market-based read on US recession risk for a client note.

What we ran: a search of Polymarket and Kalshi.

odds = valyu.search(
    "US recession 2026 or 2027",
    included_sources=["valyu/valyu-polymarket", "valyu/valyu-kalshi"],
    max_num_results=5,
)

What came back: five Polymarket markets, priced at 16:59 UTC on November 9, including:

Each record includes the full resolution rules. Both markets resolve "Yes" on two consecutive quarters of negative real GDP growth reported by the Bureau of Economic Analysis, including advance estimates, or on an NBER announcement of a recession.

What we checked at the source: the market rules and volumes in the returned records, which link to each Polymarket event page.

The catch: the 2027 market has about 1% of the 2026 market's volume, so its 28% can move on small trades. And the rules count advance GDP estimates, which can be revised, so the market tracks a definition that is not quite the NBER's.

The decision: quote the 2026 figure with confidence and the 2027 figure with its volume, and describe the resolution rule in the note.

6. Literature search: what's new on a method

The problem: a PhD student in economics is reviewing work that uses large language models as simulated participants in economic experiments.

What we ran: a search of arXiv for papers since 2025.

papers = valyu.search(
    "large language models as simulated economic agents in experiments",
    included_sources=["valyu/valyu-arxiv"],
    start_date="2025-01-01",
    max_num_results=8,
)

What came back: 8 papers from 2025 and 2026, including a study of whether generative AI agents behave like humans in laboratory market experiments (arXiv 2505.07457), work on LLM rationality in experimental economics (2505.21371), a study of LLMs replicating human cooperation in game theory experiments (2511.04500), and a multi-agent framework for economic and public policy analysis (2502.16879).

What we checked at the source: the arXiv records for the papers named above.

The catch: two of the 8 were outside economics, one simulating legal systems and one on literary studies. The search matched the method, LLM-based simulation, across fields. Screen titles and abstracts before you add them to a review.

The decision: keep the six economics papers, read their reference lists, and add Korinek's NBER paper for the methods section.

What the six jobs cost

At pay-as-you-go rates of 1,600 credits for $10, the 12.4 credits cost about $0.08.

Go further with DeepResearch: a release-day brief in one task

The jobs above each pull one set of numbers. An economist's actual output is a note: the CPI preview, the country outlook, the recession-risk update. DeepResearch writes these. It plans the research across official series, forecasts, prediction markets and the web, and returns a cited report with a CSV of the data it used.

task = valyu.deepresearch.create(
    query="Write a CPI release brief for the October 2026 US CPI: the latest official data, what markets expect, and what to watch.",
    mode="standard",
    research_strategy="Use BLS for headline and core CPI and the main components, using unadjusted series for 12-month changes and adjusted series for monthly changes. Note any missing months and how BLS treats them. Use Kalshi for market expectations with the time priced, and FRED for rates. Give the series ID and date for every figure.",
    report_format="A 600-word brief: Latest data, Market expectations, Components to watch, Risks. Every figure cited with series ID and date.",
    search={"included_sources": ["valyu/valyu-bls", "valyu/valyu-fred", "valyu/valyu-kalshi", "web"]},
    deliverables=["A CSV of every series used, with series ID, date, value and seasonal adjustment"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six jobs: say which series and adjustment each figure uses, flag missing months, and time-stamp market odds.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows include several for macro and policy work, each with set inputs and file deliverables:

To run one from code, pass its slug as workflow_id:

task = valyu.deepresearch.create(
    workflow_id="hf-macro-brief",
    workflow_params={"thesis": "US core inflation stays above 2.5% through 2027"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

For a worked example, see the macro release evidence brief and running a workflow in the docs.

Where AI agents stop in economic research

The jobs above show what an agent can do with official data. Four limits shape how to use one:

• Series choice is a judgement. Seasonally adjusted or not, headline or core, monthly or daily: the agent returns a series, and the economist decides which one answers the question.
• Data has gaps and revisions. Missing months, revised estimates and different publication schedules all affect comparisons. Record the series ID, the date and the source edition.
• Markets are evidence, not forecasts. Prediction market odds reflect traders' money and the market's rules. Read them with volume and resolution terms.
• Models and forecasts stay with you. Estimating, forecasting and interpreting the data remain the economist's work; the agent supplies the inputs.

How to set up an economic research agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp and ask the questions above in plain English. To build your own agent, see how to integrate economic data into your AI and how to build a financial research agent with Valyu.

Three habits made the jobs above reliable:

• Name the series, frequency and period. "Monthly since 2000" returned a 322-row series instead of 16,918 rows.
• Keep the metadata. Series ID, seasonal adjustment and retrieval date belong next to every figure.
• Check one figure at the source. BLS, FRED and the IMF all have public APIs, and one check catches most errors.

FAQ

What are AI agents for economic research?

AI agents for economic research are language models that search official statistics such as BLS, FRED, IMF and World Bank data, prediction markets and research papers, read what they find and return cited, structured answers. They handle the data gathering and first reading that come before an economist's analysis, as described in Anton Korinek's NBER paper "AI Agents for Economic Research".

How can AI help economists?

AI helps economists most with data retrieval and reading: pulling official series with their IDs, comparing forecasts across countries, reading prediction market odds, building charts from long series and searching new papers. In a November 2026 test, six such jobs took 8 searches and 24 seconds, and every figure matched BLS, FRED or the IMF.

Can AI agents get data from FRED?

Yes. Valyu's FRED dataset returns full series with their FRED ID, frequency, observation count, last update and vintage date. In a November 2026 test, a question asking for the monthly 10-year Treasury yield since 2000 returned series GS10 with 322 observations, and October 2026's value of 5.13% matched FRED.

What is the best AI for economic research?

AI tools for economic research fall into two groups. General assistants such as ChatGPT and Claude are good at reading and explaining. Data-connected agents reach official statistics directly. Valyu connects agents, or Claude and ChatGPT through MCP, to BLS, FRED, IMF, World Bank, Kalshi, Polymarket and arXiv through one API, with series IDs and dates attached.

What data sources do AI agents for economics need?

AI agents for economics need official statistics (BLS, FRED, BEA, national statistics offices), international data (IMF, World Bank), market data and prediction markets (Kalshi, Polymarket), and research papers (arXiv, NBER, journals). Valyu searches BLS, FRED, IMF, World Bank, Destatis, USAspending, Kalshi, Polymarket and arXiv through one API.

Can AI read prediction market odds for economic events?

Yes. In a November 2026 test, an AI agent read Kalshi's October CPI markets the day before the release: 55% for a monthly rise above 0.2% and 9% above 0.3%. It also read Polymarket's recession markets with their rules and volume, which matters because the 2027 market had traded about $22,900 against $2.4 million for 2026.

How accurate is AI for economic data?

AI is as accurate as the series it retrieves and the checks you run. In a November 2026 test, every figure matched its source, but three needed care: the search returned seasonally adjusted CPI when the headline 12-month change uses unadjusted data, October 2025 CPI is missing because of the 2025 shutdown, and the World Bank and IMF disagreed on UK growth in 2025.

Start with one job

Pick the job your team repeats most often, such as the release-day brief or the cross-country table, and run it against figures you already have. Compare the agent's numbers and series IDs with yours.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For research on companies and markets, see AI Agents for Financial Research.