The 100x Researcher: How an AI Agent Replicated a Stanford Study for $10

For the past few months, Stanford Professor Andy Hall ran an experiment that felt both thrilling and vaguely unsettling: could he automate himself?

What does a research institution look like when it operates at 100 times the scale and speed of academia today? To find out, Hall had an AI agent entirely replicate and extend his own complex, months-long study on universal vote-by-mail.

The original paper required a highly trained team, custom data cleaning in Stata, and rigorous robustness checks. It cost thousands of dollars and took months. But a single instance of Claude Code, given an autonomous terminal environment, converted the data to Python, corrected its own coding errors, and completed the analysis in under sixty minutes.

The Sandwich Metric

When evaluating the impact of AI, we often focus on the time saved. But the most staggering metric here is the cost. The total expense to run Claude Code autonomously for an hour, iterating through complex data structures and outputting verified statistical models, was about ten dollars.

The price of a sandwich.

How a highly trained team and months of custom data cleaning were replaced by an agentic swarm for the price of lunch isn’t just an optimization—it represents the commoditization of base-level execution.

The Death of the Static Whitepaper & Structural Reproducibility

The real paradigm shift here isn’t just about speed or saving money. It’s about a fundamental transformation in the structure of knowledge itself.

Currently, academic research is frozen in time. You publish a static PDF, and by the time anyone reads it, the world has moved on. The 100x researcher replaces the static whitepaper with a “Living Dashboard.” Every time new data arrives, the agent detects it, re-runs the analysis, and updates the estimates in real time.

This leads to what Hall calls Structural Reproducibility. Instead of trusting a peer reviewer to manually verify the math, the primary output of a study becomes an executable AI agent trace.

If the script fails to produce the exact same coefficients, the paper literally cannot be published. It replaces “Trust me, I did the math” with “Here is the machine that did the math; press start.”

The Flaws: Hallucinations and Nuance

Of course, this approach has notable flaws. These systems still hallucinate, misinterpret nuances in the data, and make careless integration errors.

AI models cannot invent a new natural experiment, nor can they perfectly judge which variable matters most for democracy. That still requires profound human verification. The voice and the code may be AI, but the architectural insights and moral judgments are 100% human.

The Future of Knowledge Work

Yet, the deep implication remains clear. When the cost of execution drops to zero, the nature of the PhD must change.

Young scholars must stop training to be “code monkeys” and mechanical data cleaners. Instead, they must become Architects of Inquiry—designing the prompts, the parameters, and the strategic direction that guides these vast agentic swarms.