The ways AI is transforming the news industry are coming fast, the latest being agentic AI.

During the recent INMA Agentic AI Master Class, leaders from Hindustan Times Digital, The Washington Post, Mediehuset Tromsø, and Ippen Digital shared their agentic AI strategies, experiments, and successes:

Hindustan Times Digital 

When the conversation turns to AI in media, most organisations are still talking about copilots, experimentation, and isolated pilots.

Hindustan Times Digital is talking about orchestration.

Amit Verma, chief technology officer at Hindustan Times Digital, delivered one of the clearest operational case studies of what it means to move beyond AI assistance and toward agentic systems embedded across an entire organisation.

Verma structured HT’s AI journey across three primary domains:

1. Product engineering and development.

2. Growth, marketing, and sales.

3. AI-first consumer products.

The story begins, unsurprisingly, with engineering.

Like many organisations, HT’s engineers initially embraced ChatGPT informally. Then came Copilot. But the real transformation began when HT stopped thinking about AI as a helper and started designing it as a team.

Verma described the shift clearly: “We moved to agentic AI development, where we created autonomous agents.”

Instead of individuals using AI tools independently, HT built a coordinated system of agents that mimics its engineering structure. There is an engineering manager agent that takes a Jira requirement and breaks it into subtasks. Frontend and backend agents implement. A QA agent generates and runs test cases. Human oversight remains in the loop, but the workflow is no longer dependent on linear handoffs.

The scale of impact is difficult to ignore. In HD media, “97.4% of lines of code are generated by AI,” said Verma. “In the last 28 days alone, we have generated 1.7 million lines of code.”

With the same number of engineers a year earlier, output was closer to 300,000 lines over a similar period. That represents a 6x-7x increase in code production — without increasing headcount.

Equally important, quality improved: “The number of functional bugs and production bugs has dropped by a staggering 70% or so in the past two years after the adoption of AI.”

The Washington Post

Technology is changing how audiences consume news, and that has tasked news media companies with finding new ways to make it easier and more enjoyable for them to keep up. At The Washington Post, that raised a new question: What would happen if audio journalism stopped being a one‑way broadcast and became a conversation?

Bailey Kattleman walked through how that question evolved into a new AI‑driven listening experience. The product leader for the Post explained that this new experience, which is currently still in beta, is designed to meet audiences where they are, adapt to their needs, and rebuild the habit of staying informed.

The changing audio habit: Social platforms and conversational AI tools have trained users to expect personalisation, flexibility, and interactivity — not the one-size-fits-all approach of traditional news products. Kattleman’s team built a product that adapted to the listener and expand the reach of the Post’s journalism by making it easier — and more delightful — to stay informed.

Making listening interactive: The challenge was to introduce interactivity without compromising the editorial standards that define the Post. They wanted listeners to be able to ask follow‑up questions in the moment — much like audiences used to do with call-in radio shows — without leaving the audio experience. That was important because many listening moments are hands‑free: driving, walking, cooking, getting kids ready. 

The challenges of a new product: Testing the product revealed a truth familiar to anyone building with AI: It’s not like testing a traditional product. When something sounded off or editorialised too much, engineers couldn’t always estimate how long a fix would take. Internal alignment was another hurdle. Newsrooms have extremely high standards for accuracy and tone, and the team had to define what “good enough” meant in an AI context, where perfection is impossible. That meant categorizing error types and establishing thresholds for acceptable and critical errors.

Mediehuset Tromsø

Mediehuset in Norway recently covered a big story about how the Airbnb rental explosion in Tromsø has created a housing crisis. Many of the articles on that were detected by agentic AI, shared Editor Rune Ytreberg at the INMA Agentic AI Master Class.

“Our reporters got help from a layer of agents to do the analysis and find the stories,” he said. 

The team wanted to produce not only stories about human consequences, but also an investigation into the system complexity of how and why the housing crisis is happening.

“That meant we had to go from traditional, investigative, data-driven journalism to try to redesign our newsroom using agents.”

Ytreberg introduced INMA members to LARS, the Layered Agent Research System created by Mediehuset. The system has three layers:

1. Layer 1: Evidence infrastructure agents

2. Layer 2: Analytical agents

3. Layer 3: Story intelligence and journalist collaboration

The team looked at how they should go from examining the real-life impact on people to understanding the complex system that produced these consequences.

One of the first stories produced was about Jennifer Lund, who feared she would become homeless soon because Airbnb had priced her out of home affordability.

“We saw that the housing crisis in Tromsø was so big that 140 people were living in emergency housing,” Ytreberg said. “At the same time we saw there was extreme growth in short-term Airbnb rentals. So it was a lot of emotions, a lot of people angry about Airbnb.”

The journalists wanted to find out what effect Airbnb was actually having on the housing crisis. To do that, the team built an agentic system.

“[LARS] is not a chatbot; this is not a writing assistant. This is not automation,” Ytreberg clarified. “This is a layered investigative architecture with editorial supervision.”

Human experts oversee and approve all output from the agents, as well as giving the agents tasks and subtasks. These layers worked together on research, analyses, synthesising information, and looking for stories in the data.

There is no automated publishing. Ytreberg said he’s against any automated publishing. 

“If you do that, you take out all the good human stories.” 

Ippen Digital

Markus Franz has seen the future, and he shared it with INMA members. 

The chief technology officer and incubator lab lead at Ippen Digital said he recognises that news media companies “will have some kind of support relationship with virtual teams or multi-agents.”

Built on each newsroom’s institutional knowledge, these agents will “have access to our valuable work, to our stories, to our articles” — and that will change everything.

Franz discussed three distinct chapters that looked at acceration, architecture, and retrieval.

1. The age of acceleration: News media organisations have entered “the decade of agents” — a period in which isolated AI tools will give way to autonomous, multi‑agent teams capable of researching, reasoning, and producing insights at a scale previously unimaginable.

“Machine intelligence getting faster and faster, cheaper, and smarter,” he said, noting every model in use today, is already a “legacy model” that will be surpassed in months.

The release cycle has accelerated to the point where new frontier models — GPT‑5.2, GPT‑5.3, Anthropic’s latest Claude versions, Google’s Gemini 3.1 — arrive quarterly, each bringing new reasoning capabilities and longer autonomous run times.

2. The five levels of AI: Franz traced the evolution of AI: from simple chatbots, to reasoning models, to today’s multi‑agent systems. The next step, he explained, is innovative agents — systems that not only execute tasks but generate new insights, products, and research. “These agents will research with our data and build new products or new insights.”

That has profound implications for media organisations. The question is no longer whether AI can assist with tasks, but how it will integrate into the newsroom’s workflows:

“This is what we are thinking about in the lab,” he said. “What does it mean to have this kind of system? How can this system and research system improve our products and our content and so on.”

3. Understanding the architecture: Agentic AI does away with specific models like OpenAI and Anthropic, instead leaning into agents working together to fulfill a job. They will plan stories, research them, refine them and more.  

At the heart of it all is institutional intelligence in which each newsroom’s archive, notes, transcripts, and raw data create a reservoir of knowledge that agentic systems can finally unlock.

“Every publisher has an archive,” he said. “We are producing a lot of valuable content, and we need a better way to understand and work with that content.”

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .