# AI agents are starting to do the work of research economists on Wall Street > Source: > Published: 2026-10-24 22:24:38+00:00 *A Philadelphia Fed economist built an AI agent that runs economic research on its own, and the banks cutting junior analyst classes by two-thirds are watching closely.* The people whose entire job is predicting what the economy does next are now testing whether a machine can do that job better than they can. In August 2026, Minchul Shin, a senior economic advisor and machine learning economist at the Federal Reserve Bank of Philadelphia, published a working paper that should worry anyone with "economist" in their job title. It's called "An Auditable AI Agent Loop for Empirical Economics: A Case Study in Forecast Combination," and it describes an AI agent that runs independent empirical research loops, searching for better ways to combine economic forecasts, without a human economist directing each step. Shin's paper, published as Philadelphia Fed working paper 26-41 and posted to arXiv, isn't a thought experiment. He took an open-source AI agent-loop architecture, the kind built for writing and running code, and pointed it at a real empirical economics task: improving forecast-combination methods, the statistical techniques economists use to blend multiple predictions into one. The agent searched on its own, and in several independent runs it found methods that beat the benchmarks from the original study it was testing against. That's the headline result. The caveat, which Shin built into the paper on purpose, is that not every improvement held up once he checked it against a holdout sample the agent never saw during its search. Some of what looked like genuine progress was just the agent finding a pattern that fit the specific data it had access to. That caveat is the whole point of the paper, and it's worth sitting with for a second. AI coding agents make it fast and cheap to try hundreds of model specifications, which sounds like a research superpower. But economists have a name for the risk hiding inside that speed: researcher degrees of freedom, the idea that if you try enough variations, you'll eventually find one that looks good purely by chance. Shin's fix is to log every search the agent runs and test the survivors against data held back from the start. It's a transparency mechanism, not a replacement plan. But it's also a working demonstration that an agent can do a meaningful slice of what a research economist spends their career doing: propose a model, test it, refine it, repeat. You don't have to squint to see where this is headed in finance. Goldman Sachs, JPMorgan, Citigroup and Barclays are already shrinking their junior analyst classes by as much as two-thirds, according to reporting that cites McKinsey's QuantumBlack practice, even as those same banks pull roughly 62% of their AI talent from those same entry-level cohorts. That's not a coincidence. It's banks eating their own pipeline because AI tools now do the pitch-book building and financial modeling that used to be an analyst's first two years on the job. [Anthropic's Jack Clark Tells the IMF AI Could Push US GDP to $44 Trillion](https://startupfortune.com/anthropics-jack-clark-tells-the-imf-ai-could-push-us-gdp-to-44-trillion/) Anthropic's own Economic Scenario Explorer models three 2030 paths: $34.1 trillion if AI diffuses like the internet, $36.3 trillion if it automates half of knowledge work, or $44.4 trillion in the extreme case Clark cited. - [Anthropic's Jack Clark IMF AI GDP predictions](https://startupfortune.com/anthropics-jack-clark-tells-the-imf-ai-could-push-us-gdp-to-44-trillion/) - [AI economic impact on US GDP growth forecasts](https://startupfortune.com/anthropics-jack-clark-tells-the-imf-ai-could-push-us-gdp-to-44-trillion/) Goldman Sachs president John Waldron has described his own bank, on the record, as a "human assembly line," telling CNBC that generative AI "digital agents" will serve as the bank's new "robots" on that line. Citigroup CEO Jane Fraser has told staff directly that some roles "will no longer be required." JPMorgan's Jamie Dimon said in December that AI "will eliminate jobs," full stop, no hedge. These aren't outside critics predicting disruption. They're the people running the banks, saying it themselves. Here's the thing that makes research economists a different story than junior analysts, though. Pitch books and financial models are structured, repeatable tasks: slot in the numbers, follow the template. Forecasting the economy is not that. It requires judgment about which variables matter this cycle, when a historical relationship has broken down, and when the data is simply noisy. Shin's agent found real improvements in a narrow, well-defined task, forecast combination, and still needed a human-designed holdout test to tell the real gains from the lucky ones. That's not nothing. But it's also not an agent that can watch an unexpected CPI print and tell a trading desk what the Fed does next without someone checking its work. Goldman Sachs' own economists, Joseph Briggs and Sarah Dong, estimated in a January 2026 report that AI could eventually automate close to a quarter of all US work hours, displacing somewhere between 6% and 9% of jobs and pushing unemployment up by roughly 0.6 percentage points, something like a million workers. That report was about the economy broadly. The Philadelphia Fed paper is a glimpse of the same force arriving inside economics departments themselves, including, eventually, the ones that write reports like Briggs and Dong's. The irony practically writes itself. The discipline that's spent the last few years calculating how many other people's jobs AI will take is now the one building the tool that automates its own research process. Frankly, that's not a bad place to end this story. The economists forecasting disruption and the economists getting disrupted might turn out to be the same people, just a few years apart. **Also read:** [AI neocloud stocks sink as CoreWeave debt fears spread to equities](https://startupfortune.com/ai-neocloud-stocks-sink-as-coreweave-debt-fears-spread-to-equities/) • [ARC Prize rejects OpenAI's claim that Astra scored 99.9 percent on AGI test](https://startupfortune.com/arc-prize-rejects-openais-claim-that-astra-scored-999-percent-on-agi-test/) • [How to Calculate the Right AI Agent Error Budget Before You Ship](https://startupfortune.com/how-to-calculate-the-right-ai-agent-error-budget-before-you-ship/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.