When an agent surfaces a lead, what becomes the evidence?

Northwestern’s Generative AI in the Newsroom program just named three winners of its Agentic AI Investigative Challenge. More than 60 teams entered, and the organizers reviewed about 200 Agent Skills built for Claude Code against House and Senate lobbying disclosures plus a Congress Press corpus running from 2022 through March 2026. 12

First place went to Emma Bingham and Ian Clester. Their setup spends more on a “sweep” that designs reusable SQL screens, then runs a cheaper everyday loop that filters weak candidates before an editor sees them. Promising leads land in durable bundles of hypotheses, evidence, and logs, and multiple agents check for holes and alternative explanations. 1

Second place, Muck by Ulysses Pascal, keeps a persistent document index and uses progressive disclosure so the agent starts with summaries and samples, then opens full passages only when needed. Each citation carries a document ID, character range, and hash. Third place, Registry Nako by Nataliia Kobylchak, puts overlaps and missing-disclosure checks in Python tools, uses nine narrow skills mainly to coordinate and interpret, and finishes with a verification step that recomputes every total and shows the formula. 1

The open library on GitHub publishes the winning workflows only. It does not ship findings, traces, or evidence bundles from the challenge runs. 3

The Agentic Media question sits one step upstream of the lead itself: once an agent can surface candidates cheaply, who decides what counts as a citeable trail. That reading is an interpretation of the published workflows.

Discussion: When an agent hands you a lead with citations attached, what still has to stay under human control before the story moves?