Robert Whitehead, lead of INMA’s Digital Platform Initiative, recently took a deep dive into agentic systems as part of an intensive MIT course. 

Whitehead and other media leaders discussed the shifts facing the news media industry as agentic systems reshape it, during the opening module of the Agentic AI for News Media Master Class, hosted as part of INMA’s Product & Tech Initiative.

It is not another incremental tool but as the long‑awaited catalyst for true business transformation, Whitehead said: “It’s in many ways the thing we’ve been waiting for,” adding media companies have spent years struggling to modernise. “We’ve got no further excuses and now we’ve got the armoury that we’ve been dreaming of.”

What is agentic AI?

The first of these models in OpenAI was the o1 mini that came out in September 2024, and every version since has been this reasoning which is now the default modality for large language models.

The old benchmark of GPQA — Google proof questions and answers — has become useless because the LLM models exceeded 100% of the benchmark. So a new benchmark was developed: Humanity’s Last Exam (HLM), which is an exam that no human can do. LLMs have had a relentless improvement in accuracy on this benchmark.

“Reasoning has enabled this continuing and dramatic increase in the capability of these tools,” David Caswell, founder of Storyflow, said. “So what exactly is an agent? What’s agentic AI?”

It’s basically a way to use AI to achieve a goal, he explained. The model takes the goal and breaks it down into a set of tasks that would need to be done to achieve that goal — and then actually do those tasks, such as research, API calls, emailing, etc.

“It can do what it needs to do, in terms of tasks, to reach the goal,” Caswell said. “It’s assessing things and making sure it’s done a good job, and when it’s got all that to a satisfactory point, it can bring the results back and package it up for you.”

The three elements of an AI agent are:

• Intelligence (Large Language Models, inference-time reasoning)
• Memory (Long-term persistent memory, keeping track of its work and user)
• Tools (Search, email, APIs, payments, etc.)

Caswell also shared the characteristics of an AI agent:

• Autonomous (give it a goal and let it figure out how to achieve it)
• Asynchronous interface (give it instructions then go away)

A new product culture

Whitehead introduced the concept of “vibe coding” — using natural‑language prompts to build functioning product prototypes in minutes.

“It isn’t the future of media production, but it is potentially the future of how we think about products and how … the majority think about business development.

He demonstrated how he created a working news‑video app in just seven minutes using a semantic prompt and a no‑code platform. The app aggregated 20 short news clips from North American broadcasters, ending with a playful confetti animation and a message telling users they were “all caught up.”

The exercise illustrated both the promise and the pitfalls of vibe coding. It’s easy to build a prototype without aid from the tech team, and it greatly lowers the cost of experimentation and cross‑functional collaboration.

However, it’s not without risks, including hidden security flaws, unclear accountability, and a “false sense of readiness” when prototypes look polished but are not production‑ready. “It does erode engineering judgement,” he cautioned, because teams may assume a prototype can be deployed without proper rebuilding.

“It has lots of upside, but the downsides are real,” he said.

Rethinking the business model

Florent Daudens, CEO and co-founder of AI start-up Mizal, said prior to the prevalence of AI, automation, and bots, the premise of an online business model for news involved an initial Web search which generated search results.

From there a user or consumer would take a highly important step within the business model — a click on a piece of content. Thanks to AI, this crucial step of clicking is being erased.

Daudens believes the foundations of monetisation fit into two ways: training and inference.

“Training is like selling the rights back to your catalog,” Daudens said, while inference has the potential to drive real revenue. Training would involve ongoing maintenance to keep information accurate for content modules.

Daudens referenced an OpenAI blog post about the number of queries for local news reaching one million per week. Questions like, “What’s happening around me?” allows a user or reader to localise their news and control the type of content they want to receive. The future of journalism will likely need to consider a query-first model or news content driven by queries.

To be successful within an agentic AI world, Daudens suggested that publishers need to understand the landscape, including what content to block or not. Comprehension of Web architecture is also important.

“You really need to declare your terms, rights and permissions, access and enforcement prior to planning to accept payment and any monetisation,” Daudens said.

The monetisation of news content forces publishers to rethink not only the type of content they create but how the content is produced for both humans and agentic as master class moderator Jodie Hopperton pointed out during Dauden’s presentation.

In response, he said, “I think you either need to be really, really clear about your value proposition in terms of depth of your coverage, the uniqueness of your voice and your brand; or you pull on distribution at scale.”

Many news organisations sit in-between these directions, Daudens said, which puts them in a “very, very precarious position.”

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