From Coder to Conductor: How Generative AI and Agentic Workflows Revolutionized Data Journalism in Italia Fragile

By Jacopo Ottaviani

Executive Summary & Main Facts

In the evolving landscape of digital media, generative artificial intelligence has moved beyond basic text generation to fundamentally rewrite the production pipeline of data journalism. Writing zero lines of code manually, data journalist Jacopo Ottaviani recently conceptualized, built, and deployed Italia Fragile—a fully interactive, searchable geovisualization atlas mapping hydrogeological vulnerability across all 7,900+ Italian municipalities.

Utilizing advanced frontier AI models (specifically Anthropic’s Claude and its experimental coding environment, Claude Code), Ottaviani completed in roughly 10 iterations over a few hours what traditionally would have taken a specialized engineering team weeks or months.

However, the core takeaway is not merely speed; it is structural transformation. By leveraging agentic AI commands—specifically /goal and /loop—the author transitioned from a hands-on software builder to an editorial orchestra conductor. In this paradigm, artificial intelligence handles data scraping, statistical joining, UI prototyping, and ongoing maintenance, while the human journalist retains ultimate editorial oversight, defining the initial questions, stress-testing the ethics, and maintaining accountability.

The Genesis of Italia Fragile: Chronology and Context

From Static Archives to Agentic Infrastructure

The roots of this technological pivot trace back to earlier data journalism projects rebuilt by Ottaviani using AI: Patrie Galere (a map of Italian prison fatalities originally developed in 2012) and Strade Mortali (a portal tracking road accidents in Rome assembled during a rapid two-day sprint). While those earlier experiments proved that large language models (LLMs) could write working code, they still relied on traditional, one-off conversational prompting.

With Italia Fragile, the workflow evolved into agentic computing. The project was built using Fable 5, an advanced frontier model. (Access to the model was briefly interrupted in June by US authorities for security reviews regarding foreign nationals, before being restored).

The development timeline moved at unprecedented speed:

1. Ideation & Sparring (Hour 1): Using natural language to define the conceptual boundaries of the atlas.
2. Research & Textual Ingestion (Hour 2): Ingesting and summarizing hundreds of pages of dense technical environmental reports.
3. Design & Prototyping (Hours 3–5): Instantiating interactive mapping components via LeafletJS through natural language refinement.
4. Data Acquisition & Harmonization (Hours 6–8): Automating complex web scrapers and joining disparate administrative registries.
5. Scheduled Maintenance (Ongoing): Establishing self-updating background loops to ensure the atlas functions as a living observatory rather than a static publication.

Supporting Data: The Geological Reality of Italy

Italia Fragile was not built in a vacuum; it responds to an alarming national crisis documented extensively by state institutions. According to the 2024 report on hydrogeological instability published by the Italian Institute for Environmental Protection and Research (ISPRA):

• 94.5% of Italian municipalities are classified as being at risk from landslides, floods, coastal erosion, or avalanches.
• 5.7 million people reside in high-hazard landslide zones, with 1.28 million of them concentrated in the two highest severity brackets.
• Over 636,000 individual landslides have been officially mapped nationwide, covering nearly a quarter of Italy’s entire landmass—a 15% expansion since 2021.
• Escalating Costs: National expenditure on repairing structural damage from floods and landslides has tripled since 2009, surging from a post-war average of roughly €1 billion annually to €3.3 billion, according to the national builders’ association ANCE and research center CRESME.
• The Prevention Paradox: While a dedicated national register tracks 25,539 publicly funded prevention projects worth €19.2 billion initiated since 1999, only 27% of those funds correspond to completed works. Long-term infrastructural investment continues to lag far behind the perpetual, costly cycle of emergency management.

Step-by-Step Breakdown: How AI Powered the Five Stages

Ottaviani details how agentic tools altered every phase of the investigative and developmental workflow:

1. Ideation and Conceptualization

Data journalism begins with core hypotheses. Using the /goal command, Ottaviani initiated the project with a standing directive:

"/goal make a mobile-friendly, interactive open data site (aka atlas) based on ISPRA and ISTAT datasets to map and visualise geological risk in Italy…"

A secondary goal stress-tested the concept to ensure clear distinctions between hazard (probability of an event), exposure (people and structures in harm’s way), and risk (the combination of both). While the model excelled at listing comparable projects and flagging edge cases (such as high-hazard zones with zero population), it could not replace human editorial judgment regarding newsworthiness.

2. Research and Analytical Rigor

ISPRA’s 2024 technical reports are dense with regulatory nomenclature (such as P3 and P4 classifications for extreme landslide danger). A single goal prompt processed these documents in under an hour, explaining classification systems in plain language and tracking longitudinal changes.

To prevent hallucinations, a strict verification rule was embedded:

"/goal Every figure displayed on the site must trace back to a record of our source files; keep checking until none is left untracked."

The AI audited its own analytical outputs against raw data cells, allowing the journalist to focus on interrogating the underlying policy failures—such as why three-quarters of prevention funding remains unspent.

3. Design and Prototyping

Traditionally the most capital-intensive phase, interface creation was compressed into hours. Utilizing LeafletJS, the model autonomously established color palettes, typography, and choropleth mapping schemes. When provided with subjective feedback—such as "the risk colours read as decorative rather than serious"—the agent immediately adjusted its CSS and design tokens. Because prototyping is practically free and instantaneous, creators can build, test, and discard multiple design iterations.

4. Data Acquisition and Scraping

Unifying data from ISPRA’s IdroGEO platform and the Italian National Statistics Institute (ISTAT) presents a notorious administrative trap: municipal boundaries shift, merge, and get renamed over time, corrupting standard database joins. The LLM wrote custom scraping scripts and successfully resolved administrative drift by recognizing near-matches and querying historical boundary changes, bypassing hours of manual spreadsheet cleaning.

5. Maintenance and Automated Living Observatories

Journalism has historically struggled with project decay—digital pieces launch, funding ceases, and links rot. Italia Fragile solves this via automated scheduling commands:

"/schedule monthly: check IdroGEO for new data editions; if found, re-run the scrapers and joins, rebuild the atlas and report what changed."

By combining language models with cron-like scheduling, the project transitions from a static investigative report into a permanent, self-updating public observatory.

Official Responses and Institutional Implications

The deployment of automated, transparent environmental tools highlights a growing tension between public data availability and institutional transparency. While organizations like ISPRA and ISTAT publish vast repositories of open data, local governments and regional authorities have historically struggled to communicate these risks effectively to constituents.

Environmental safety advocates have welcomed user-friendly prototypes like Italia Fragile, noting that democratizing access to complex hazard indicators empowers local councillors, grassroots activists, and regional journalists to hold policymakers accountable. Conversely, public works agencies face mounting scrutiny over the vast discrepancies between allocated disaster-prevention funds and actual completed infrastructural improvements.

The Future of Newsrooms: From Loops to Graphs

In evaluating the broader industry shift, software architect Boris Cherny’s framework on the Steps of AI Adoption provides a useful roadmap for newsrooms. As organizations move from banning AI to managing dozens—and eventually hundreds—of concurrent agents, the core challenge is not technological capability, but the implementation of rigorous validation checks.

Without strict protocols—such as automated record-count verifications, source-traceability rules, and mandatory human sign-offs—newsrooms risk scaling errors rather than efficiency, leading to public failures that damage institutional trust.

Beyond simple linear "loops" (single agents pursuing a single goal), advanced workflows are moving toward graph engineering. In a graph-based newsroom architecture:

• Nodes represent specialized units of work (scrapers, translators, fact-checkers).
• Edges define the conditional rules governing workflow transitions.
• State preserves context passed seamlessly between editorial steps.

Conclusion: The Conductor’s Burden

Ultimately, generative artificial intelligence and agentic workflows do not eliminate the journalist; they elevate them. While an AI agent can tirelessly execute code, scrape databases, and verify records against constraints, it possesses no intrinsic understanding of human values, community impact, or moral urgency.

In the era of automated data journalism, the writer’s primary role has evolved from a lone builder typing lines of code into an orchestra conductor—coordinating specialized digital agents, setting the artistic and ethical tempo, and ensuring that the final performance resonates with the public it serves.