Coding agents in the social sciences

The human sciences are shifting: for the first time, core research tasks can be handed off to machines. AI increasingly contributes to research, including in the most prestigious publications and in the social sciences.

But while turn-taking chatbots have mostly been used for writing assistance, coding agents could restructure social-science research more radically. An agentic workbench can take a research idea and a dataset, write and run an analysis, interpret the output, and iterate — the irreducibly human steps of empirical work, now shared with a machine.

The tools exist; most researchers haven’t adopted them

Capability has run ahead of adoption. In a recent survey of quantitative social scientists, a large majority had tried AI in general, but only a fraction had folded a coding agent into their actual workflow. The gap between “I’ve used a chatbot” and “an agent runs my analysis” is still wide.

Adoption is highly uneven

The unevenness is the striking part. Adoption skews sharply by discipline (economists far ahead of many other fields), by career stage (doctoral students and postdocs ahead of tenured faculty), by institution (better-resourced departments ahead), and by demographic group. The disparities in agent adoption are larger than those in general AI use — which means a tool that could level the playing field might, without care, tilt it further.

Researchers use agents to code, not to write

Contrary to the fear that AI would automate the prose, the dominant use is generating and editing analysis code. Only a minority draft manuscript text with AI. This matches how MegaBrain Science is built: the agent lives next to a live kernel, iterating on code and data — and a reviewer checks the numbers and citations — rather than ghost-writing conclusions.

More projects started — not yet more submitted

Early signals show adopters starting more projects and posting more working papers and grant proposals, without (yet) a matching surge in journal submissions. The most likely reading: agents are excellent at the setup — wrangling, first-pass analysis, scaffolding — while the “last mile” of polishing a paper still takes human time. Timeline effects may also mean the submissions are simply still in flight.

Optimism about output, caution about the field

Researchers are broadly optimistic that these tools raise their own productivity, but markedly less sure they improve social science overall. The worries are real: congestion from a flood of papers, amplified selective reporting, and a drift toward safe, incremental work. Individual benefit, field-level risk.

Where we net out

It’s early, and these patterns are descriptive — we can’t yet separate genuine productivity gains from the fact that early adopters differ from everyone else. But the direction is clear enough to design around. That’s why MegaBrain Science leans on reproducibility and an independent reviewer: if agents are going to do more of the analysis, the results they produce have to be traceable, checkable, and hard to fool. Speed without verification isn’t progress.

This post adapts themes from published research on AI coding agents in the social sciences; the statistics referenced come from that external survey work, not from MegaBrain.