Acceleration and Its Discontents: Coding Agents Enter the Social Sciences
Anthropic surveyed 1,260 quantitative social scientists (Feb-Mar 2026) to track how "coding agents"—which, unlike chatbots, run, interpret, and iterate on analyses themselves—are reshaping research. Here are five insights, read through the lens of a developer who lives in Claude Code every day.
Insight 1: Agents aren't chatbots — 81% have tried, 20% actually use
Only 20% of respondents use coding agents regularly (more than weekly), even though 81% have tried AI chatbots for research. Among coding agent users, 86% use Claude Code.
Having tried it isn't using it. The 81→20 gap isn't capability—it's the trust hurdle of handing an agent your whole execution loop.
Insight 2: Adoption is strikingly uneven — and steeper than for AI in general
Researchers with typically male names use coding agents at more than twice the rate of those with female names, and top-25 university researchers are 40% more likely to adopt. Doctoral students and postdocs use them weekly at just over 25%; tenured professors at less than half that rate.
In Korea, agent-style tools also seem concentrated among early adopters in the dev and research community (anecdotal). Juniors out-adopting tenured faculty 2-to-1 is a rare structural edge for early-career folks—I'd grab this window while it's open.
Insight 3: Agents are used for code, not prose
97% of coding agent users (versus 77% of other AI users) use them to generate analysis code. Editing prose comes second; drafting prose is a minority practice outside economics and management.
Agents shine on code over prose because code runs and verifies itself. Point AI at work that has a feedback loop; the rest stays human.
Insight 4: Acceleration up front, a stall at the end
Coding agent users start about 25% more projects and post about 50% more working papers. But there is no evidence that coding agent users submit more new papers to journals. (The comparison is descriptive; selection effects can't be ruled out.)
The easier starting gets, the more half-finished work piles up. +50% working papers but flat journal submissions—completion is still what gets judged.
Insight 5: Optimistic about personal output (88%), skeptical about the field (70% gap)
88% are optimistic about AI's contribution to paper productivity (above the midpoint on a 1-10 scale), but 70% are more optimistic about their own paper productivity than about the field's overall impact.
Users are more optimistic because they've actually felt the upside. Maybe the field-level gloom is mostly the vagueness of people who haven't tried it.
Coding agents clearly pulled the start of research forward. Whether that acceleration carries through to completion—and to the field's health—is something the data can't answer yet. As someone holding the tool every day, my bet is that the difference shows up at the finish, not the start.
