--- title: How BBC Eye Used AI Agents to Expose Russia's Vigilante Networks description: BBC Eye built a multi-agent AI system called Haystack to sift 10,000 Russian social posts, revealing a vigilante group's anti-migrant campaign in detail. image: https://www.augmentednewsroom.org/hubfs/AI%20Icons%20Phone.png --- ![27 Outline Logo](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/27%20Outline%20Logo.svg) - [Home](https://www.augmentednewsroom.org) - [Case Studies](https://www.augmentednewsroom.org/case-studies) - [About](https://www.augmentednewsroom.org/about-augmented-newsroom) - [Founder](https://www.augmentednewsroom.org/founder-augmented-newsroom) [contact us](mailto:malito:alib@augmentednewsroom.org) ## Inside **BBC Eye's AI Agent** System ![](//443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/raw_assets/public/@marketplace/if_else_agency/Aesthetics/assets/images/icons/arrow-down.svg) ![BBC Eye Final](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/BBC%20Eye%20Final.png) **At a Glance** - **Organization:** BBC Eye (BBC), with computational journalists from Stanford University - **Country:** United Kingdom / Russia (subject of investigation) - **AI tool:** Haystack — a multi-agent AI system built on LangGraph, using LLMs such as OpenAI's GPT and Google's Gemma --- ### The Problem Investigative journalists run into the same wall again and again: there's more evidence than any one person can work through. For BBC Eye — the unit behind the broadcaster's original documentary investigations — that wall showed up as tens of thousands of social media posts tied to Russia's nationalist vigilante movement, more than any human researcher could realistically read, let alone analyze at scale. Using conventional reporting and data journalism, BBC Eye had identified Russkaya Obshina as the largest of these groups — running a campaign against migrants reportedly working in concert with Russian authorities.  What the team didn't have was a way to see the scale of the movement across the flood of social content it was generating. ### What They Built ![](https://reutersinstitute.politics.ox.ac.uk/sites/default/files/inline-images/image_65.png) BBC Eye's open-source investigators, working with computational journalists from Stanford University, built a multi-agent AI system called Haystack — their first attempt at this kind of multi-agent AI work. It was designed to mirror several jobs an investigations team normally splits among people: pulling data from a website or API, reading posts for leads, and turning raw numbers into findings. Haystack is built on LangGraph, a library for chaining together AI agents into a single workflow, capable of running on models like OpenAI's GPT or Google's Gemma. Functionally, it does three things: fetches posts from Russian social platforms where nationalist groups are active, assesses text and images for nationalist, racist, or anti-immigrant signals, and answers natural-language questions about the resulting data. ### How It Actually Worked The team describes using Haystack as feeling like a choose-your-own-adventure book: a reporter interacts with the system and moves down one of several defined paths, guided by a "supervisor" agent that routes tasks to the right place. Getting that supervisor agent to reliably follow a journalist's actual direction was an early sticking point, addressed by giving it examples of correct behavior and having it ask clarifying questions when a reporter's input was ambiguous. That's a deliberate design choice, not an oversight. The team considered building something more autonomous — feed the system one complex instruction and let it collect, assess, and analyze data on its own — and chose not to. Requiring a reporter to weigh in at each stage cut down on the model guessing its way into the wrong territory, and kept a journalist in the loop over what the AI was actually deciding. In practice, that meant building a seed list of more than ten nationalist and far-right groups, using a collection agent to gather their posts, and continually checking the system's output as the project went on. One concrete lesson from that process: asking the model for a graded judgment ("definitely," "probably," "probably not") produced messier results than asking for a simple "yes," "no," or "not sure." Underneath it all, the system was building a database the team could review — not just the AI's answers, but its reasoning for each one. ### The Result Haystack pulled roughly 10,000 posts from more than ten Russian nationalist groups and generated 55,000 separate assessments of that content, flagging signals like nationalist ideology, references to migrants, anti-migrant raids, and language describing violence against minorities. From that analysis, Russkaya Obshina appeared to have the most active on-the-ground presence of the groups studied — patrols in Russian towns and cities, and raids on workplaces, shops, nightclubs, and hostels, according to the team's reading of the data. ### Lesson for Journalists The most transferable idea here isn't the specific AI model or code library — it's the design choice to keep a reporter making decisions at every step, rather than handing the system one instruction and letting it run. The team explicitly tested a more autonomous version and rejected it, because it made the model more likely to guess and drift off course. For newsrooms considering similar tools, that's a concrete, testable design principle rather than a vague caution about "keeping humans in the loop." ### Risks and Limits The team's own estimate of time saved compared to manual review is exactly that — an estimate, not a controlled comparison. And everything Haystack surfaced was still checked by a person before it fed into the finished investigation; the system found leads, it didn't file the story. ### Sources - [https://reutersinstitute.politics.ox.ac.uk/news/how-bbc-eye-built-multi-agent-ai-system-sift-through-ten-thousand-russian-social-media-posts](https://reutersinstitute.politics.ox.ac.uk/news/how-bbc-eye-built-multi-agent-ai-system-sift-through-ten-thousand-russian-social-media-posts) ![Ali Bahnasawy](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/Ali%20Photos/Square%20Close%20up.png) [**Ali Bahnasawy**](https://www.augmentednewsroom.org/case-studies/author/ali-bahnasawy) #### My awesome resource Vivamus suscipit tortor eget felis porttitor volutpat. Curabitur arcu erat, accumsan id imperdiet et, porttitor at sem.  Discover ## Related **blog posts** ![](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/BAF%20GOOD%20EXCELLENT-1.png) **From Bad to Excellent: How Small Changes Transform Your AI Prompts** Your AI isn't the problem. Your prompt is. Read more ![](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/pexels-mizunokozuki-12899156.jpg) **Prompt Engineering for Media Professionals - Part 1** Most people get mediocre results from AI because they treat prompts like casual questions and accept the first answer... Read more ![](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/WP%20Images-1.png) **Ask The Post AI: Grounding Chatbot Answers in Verified Reporting** At a Glance Organization: The Washington Post Country: United States AI tool: Ask The Post AI — a generative AI... Read more ![Grey Augmented Newsroom](https://443642032.fs1.hubspotusercontent-ap1.net/hubfs/443642032/Grey%20Augmented%20Newsroom.svg) Augmented Newsroom helps journalists and newsrooms use AI responsibly — faster reporting without sacrificing accuracy or editorial standards. - [Home](https://www.augmentednewsroom.org) - [Case Studies](https://www.augmentednewsroom.org/case-studies) - [About](https://www.augmentednewsroom.org/about-augmented-newsroom) - [Founder](https://www.augmentednewsroom.org/founder-augmented-newsroom) ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Ali Bahnasawy", "url" : "https://www.augmentednewsroom.org/case-studies/author/ali-bahnasawy" }, "dateModified" : "2026-09-15T00:22:38.434Z", "datePublished" : "2026-09-08T04:18:08.000Z", "headline" : "How BBC Eye Used AI Agents to Expose Russia's Vigilante Networks", "image" : [ "https://www.augmentednewsroom.org/hubfs/AI%20Icons%20Phone.png" ], "mainEntityOfPage" : { "@id" : "https://www.augmentednewsroom.org/case-studies/bbceye-multi-agent-ai", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject" } } } ```