Conducting the Code: How Generative AI and Autonomous Agents Are Redefining Data Journalism

By Global Newsroom Insights
Published: October 2026

Main Facts: The Orchestration of "Italia Fragile"

In the summer of 2026, data journalist Jacopo Ottaviani accomplished what once would have taken a dedicated engineering team weeks or months: he built a fully functional, highly detailed open-data mapping portal named Italia Fragile. The interactive web application allows users to look up any municipality in Italy and immediately access official state data regarding the geological stability—or fragility—of the ground beneath it.

The core breakthrough of this project was not merely the utility of the atlas, but its creation method. Ottaviani wrote zero lines of traditional code. Instead, utilizing frontier artificial intelligence models—specifically Anthropic’s Claude code agents equipped with advanced systemic reasoning frameworks—he functioned less as a hands-on developer and more as an orchestra conductor.

Through the strategic use of standing objectives (/goal) and scheduled recurring tasks (/loop), the AI model independently handled ideation stress-testing, document comprehension, web scraping, data joins, front-end interface design, and automated maintenance protocols. This paradigm shift proves that modern large language models (LLMs) can reliably transition from conversational chatbots into autonomous agents capable of end-to-end execution in complex journalistic workflows.

Chronology: From Idea to Autonomous Observatory

The realization of Italia Fragile marks an evolution from Ottaviani’s previous experiments with AI-assisted journalism, such as rebuilding Patrie Galere (a map of Italian prison deaths) and Strade Mortali (a Rome road accident portal). The chronology of building the hydrogeological risk atlas unfolded across several distinct phases:

• June 2026: Ottaviani initially tests frontier models like "Fable" over a two-day period before temporary US security suspensions restrict foreign national access.
• Access Restoration & Prompting Phase: Upon access restoration, Ottaviani initiates the project by inputting standing goals into Claude Code rather than one-off conversational prompts.
• The Development Sprint: Over roughly ten total iterations across a few hours, the AI processes hundreds of pages of technical documentation, structures databases, writes modular JavaScript using LeafletJS, and deploys interactive map layers.
• Deployment and Scheduling: Rather than treating the project as a static, one-off publication, Ottaviani programs scheduled agents using /loop commands to continuously monitor source repositories for future data updates, effectively transforming the news product into a permanent digital observatory.

Supporting Data: Italy’s Mounting Hydrogeological Crisis

The urgency behind Italia Fragile is underscored by alarming environmental and financial statistics compiled by state agencies. According to the 2024 report by the Italian National Institute for Environmental Protection and Research (ISPRA):

• 94.5% of all Italian municipalities are officially at risk from landslides, floods, coastal erosion, or avalanches.
• 5.7 million people live in areas vulnerable to landslides, with 1.28 million residing in the two highest hazard categories.
• Over 636,000 mapped landslides exist across the peninsula, placing Italy among the highest counts in Europe and covering nearly a quarter of the national territory (a 15% increase since 2021).
• The Cost of Inaction: Repairing flood and landslide damage has tripled since 2009, jumping from a post-war average of roughly €1 billion annually to €3.3 billion, according to the builders’ association ANCE and research center CRESME.
• The Prevention Gap: While a dedicated register records 25,539 state-funded prevention projects worth €19.2 billion initiated since 1999, only 27% of those funds correspond to projects that have actually reached completion.

Official Responses and Institutional Context

The deployment of autonomous AI tools to interrogate massive government datasets has drawn significant attention from both the tech industry and public policy sectors. Boris Cherny, creator of Anthropic’s Claude Code tool, and his engineering colleagues have argued that "loops" represent the definitive next evolutionary step after basic prompting. By allowing AI systems to act, verify their own work, and iterate without constant human hand-holding, these agentic frameworks remove technical bottlenecks from specialized investigations.

Simultaneously, public institutions are confronting the implications of democratized data access. Reports produced by bodies like ISPRA have historically remained trapped behind dense technical jargon, complex classification systems (such as P3 and P4 landslide danger tiers), and multi-hundred-page PDF reports.

When autonomous AI models are directed to ingest, decode, and simplify these documents in minutes, local councillors, regional journalists, and ordinary citizens gain unprecedented analytical capabilities. This levels the playing field, allowing local stakeholders to cross-examine state figures on municipal risk without needing advanced degrees in geographical information systems (GIS) or data science.

Implications: The Conductor’s Dilemma and the Future of Newsrooms

The rise of agentic AI workflows introduces profound implications for the media industry, challenging traditional definitions of reporting, verification, and newsroom economics.

1. From Static Artifacts to Live Services

Historically, investigative data journalism projects suffered from a fatal flaw: maintenance failure. Funded by grants designed for launch rather than longevity, interactive maps frequently rotted and died when hosting services expired or datasets drifted. By embedding scheduled loops into the pipeline, journalists can now build self-updating observatories. The project ceases to be a static article tied to a breaking news cycle and becomes an enduring public service.

2. The Economics of Rapid Prototyping

With subscription costs for advanced coding tools running at modest monthly fees (such as roughly £15 for Claude Pro), the financial barrier to entry for complex web development has vanished. When building a functional, database-backed interactive map takes hours instead of months, data desks are no longer locked into their initial brainstorming ideas. They can build, discard, and test multiple variations before selecting the most effective editorial approach.

3. Navigating the Adoption Ladder

Software frameworks like those outlined by Boris Cherny suggest that newsrooms will inevitably move through stages of AI adoption—from outright bans to managing single agents, dozens, and eventually hundreds of concurrent processes. However, this ascent carries inherent dangers. Without rigorous internal verification rules—such as mandating that every displayed metric must trace back to a verifiable source cell—newsrooms scaling up agentic workflows risk mass-producing unverified errors that severely damage public trust.

4. Preserving the Human Core

Ultimately, the transition of journalists into "orchestra conductors" highlights what AI cannot do. An autonomous loop can relentlessly pursue a mechanical goal, write clean code, and cross-reference database joins, but it cannot determine whether a dataset deserves public attention, whether an investigation serves the public interest, or what ethical weight a story carries.

As Italia Fragile demonstrates, the heavy lifting of code and computation can be safely delegated to artificial intelligence. Yet the foundational responsibilities of editorial judgment, moral clarity, and accountability remain firmly in human hands.