Cole Medin uses an Anthropic study as a starting point for thinking about judgment in coding-agent planning.

What the study tested

  • Fifty professional analysts used an agent on a fictional farming-cooperative analysis. The agent was secretly instructed to argue for randomly assigned positions.
  • Across four analytical choices, their decisions matched the agent’s assigned lean 86% of the time.
  • In 38 cases where a preference was stated first, analysts switched 60% of the time when the agent disagreed, versus 6% when it agreed; the small sample makes that comparison imprecise.
  • Afterward, 77% repeated an agent argument at least once and 15% credited the agent anywhere in their answers.
  • Caveat: The researchers deliberately strengthened the agent’s lean, and studied one task in one field with one model. This does not establish those rates for coding agents.

Keep a hand on the wheel

  • State priorities first: Describe the problem, pain and constraints before asking for recommended solutions.
  • Prototype alternatives: Build disposable versions when a decision is hard to visualize, then compare them against your priorities.
  • Ask what would change the advice: Request the conditions under which the other option would be better, and check whether they apply.
  • These are Medin’s proposed practices, not interventions tested by the study.

Independence and oversight

  • The study raises a concern that apparent independent reviews may share the same model’s assumptions.
  • AI-assisted analysts made fewer serious errors and finished faster in this experiment, but the small comparison group chose not to use AI and was less experienced; causation is uncertain.
  • Medin’s sponsored payment example relies on API permissions that permit drafts but not sending payments, followed by human approval.

“What would have to change … so option B is optimal?” — Cole Medin