Product

Reporters Workbench AI

an agent‑augmented “notebook” that helps journalists turn raw notes into actionable tasks and evidence‑based insights

Reporter’s Workbench AI is a desktop web app that lets a journalist group notes into story projects, converts fresh notes into suggested tasks via OpenAI’s o3 function‑calling capabilities, and automatically executes a small set of high‑value tools within those tasks—web search, long‑PDF search and extraction, and quick dataset queries—using OpenAI’s search tool, PDF parsing, and DuckDB. By measuring end‑to‑end task‑completion time, the prototype will test whether a lean, agentic workflow can meaningfully accelerate early‑stage reporting without sacrificing control or transparency.​

Faculty and Staff Leads

Nick Hagar

Postdoctoral Scholar, Generative AI in the Newsroom Initiative

Nick is a postdoc developing genAI applications for newsrooms. They research how collective attention works in large, complex systems: how people discover information online, why things get popular, and what influences content creators. Formerly at the New York Times, Meta, and others.

Project Details

2025 Fall

Important Questions
  • Which agentic GenAI tools most effectively accelerate the early‑stage reporting tasks we’re targeting (background search, long‑doc triage, quick‑look data analysis)?
  • How should responsibility be divided among fully automated, collaborative, and human‑only actions so that speed gains never outstrip editorial control or verification?
  • What safeguards—e.g., local processing, provenance‐rich citations, mandatory human‑approval gates—are required to manage hallucination risk and protect sensitive source material?
  • If the prototype proves useful, what ripple effects might it have on workflows, roles and resource constraints in small or local newsrooms?
Sample Milestones
  • Weeks 1-3: User interviews, scope lock, start building backend
  • Weeks 4-6: Begin front-end UI and task-builder agent; first usre testing
  • Weeks 7-8: Build out web search and document processing systems
  • Weeks 9-10: More testing, public-facing demo, and write-up
Outcome

By Week 10, we expect a working single‑user desktop web prototype where reporters can drop a new note into a Project, watch the Task‑Builder agent suggest actionable steps, and see at least one of those steps (web backgrounding, PDF summarization, or data analysis) execute automatically, posting its results back as a traceable Note—demonstrating a measurable reduction in task‑completion time and laying the groundwork for richer agentic workflows in future quarters.

Links