Let’s be honest: on the surface, the AI boom feels at odds with the values of traditional publishing. We are watching social platforms fill up with AI-generated ‘slop’, and we know that allowing that kind of noise onto our own platforms would be fatal. In this industry, trust is our only currency, and retaining it means being radically transparent about how and where we use machine intelligence. Thankfully, however, we have moved past the initial shock and the messy experimentation phase. We are now arriving at something that finally resembles industrial maturity.
To frame the technology landscape for digital news media as we scan the horizon of 2026, the simplest way to view it is through the relationship we have with the AI machine: in 2024, AI was a chatbot; in 2025, AI was a copilot and in 2026, AI is a colleague.
In 2024, the emphasis was all about the prompt. We treated AI like a magic 8-ball or a very fast librarian. We asked a question, it gave an answer, and we hoped it wasn’t hallucinating. In 2025, we integrated these tools into our sidebars. They became ‘copilots’ that could summarise an email or suggest a headline, helpful but ultimately passive. They waited for us to drive.
The horizon for 2026 is different. We are moving into the era of the ‘agentic harness’. This is where we stop treating AI as a tool we pick up and put down and start treating it as an active participant in our workflows – a junior colleague that can manage complicated, multi-step tasks with autonomy, provided we build the right guardrails around it.
Beyond the prompt: building the agentic harness
The single biggest shift for technology leaders in 2026 is the move away from the ‘chat interface’. The idea that a single prompt can solve a complex business problem was a nice dream, but it turned out to be a fallacy. Real work is rarely linear. It involves checking, refining, cross-referencing, and formatting. A single prompt cannot consistently do that.
Companies that want to become truly AI-native need to stop using chatbots and start building agentic harnesses. A harness functions as more than an AI model. It is a software wrapper, often built with orchestration frameworks such as LangChain or AutoGen, that breaks a job down into a chain of thought. It forces the AI to show its working.
Take a standard internal revenue report as a case in point. In most companies, this is a monthly or weekly headache. It can involve a subscription manager pulling churn data, an ad ops specialist pulling yield data, and a finance analyst trying to combine everything into a summary for the exec board.
If you paste all that data into a standard LLM and ask for a report, you will likely get a hallucination or inaccuracies at best. The numbers probably won’t add up because LLMs are language models, not calculators. It is unlikely any C-suite executive would trust such a report, nor should they.
However, an agentic harness takes a multi-step approach that mimics a human workflow:
- Agent A (The Retriever) is tasked solely with accessing your finance systems’ API to pull the raw CSVs.
- Agent B (The Analyst) does not use the LLM for math. Instead, it writes and executes a Python script to calculate the variances. This ensures 100% mathematical accuracy.
- Agent C (The Critic) reviews the output of Agent B against historical norms to flag anomalies.
- Agent D (The Writer) takes the verified data and drafts the narrative summary.
This is the difference between a gamble and a workflow. The ‘harness’ is the code that binds these agents together and, crucially, inserts validation gates between them. If Agent C spots an anomaly, it doesn’t hallucinate an excuse. It halts the process and either restarts Agent A and B or pings a human.
Andrew Ng, a leading voice in AI, has been vocal about this shift, noting that ‘agentic workflows’ often yield better results with smaller models than single-shot prompts do with massive models. By 2027, building these harnesses will likely be a large part of the media technology team’s work.
The fork in the road: compounding vs stalling
We are currently standing at a fork in the road. Your approach to this agentic shift will determine whether the value of AI in your company compounds or stalls.
The ‘stall’ happens when companies restrict AI to a licensed tool, essentially giving everyone a generic ‘enterprise copilot’ account while crossing their fingers. This leads to a plateau. You get a slight productivity bump in email writing, but you don’t fundamentally change the economics of your newsroom.
The ‘compound’ effect happens when you merge the skill trees of your technical and non-technical roles to build the harnesses described above.
In 2026, we are seeing the rise of the ‘AI operations lead’, a hybrid role that sits physically in the newsroom but operates with a product manager’s mindset. Leaders need to actively facilitate this convergence. You cannot expect your developers to intuitively understand the nuance of a ‘lede’, nor can you expect editors to grasp the mechanics of a ‘vector database’ without support. While those are extreme examples, the reality is that these skill trees are merging, and your organisational structures need to reflect it. The goal is to have an editorial team that knows how to ‘manage’ a suite of AI agents just as effectively as they manage a team of freelancers.
At The Independent, we have operationalised this by adhering to a hard and fast rule: no AI-generated content is published without a human-in-the-loop to edit and rewrite. Our ‘Bulletin’ product, launched in 2025, is the practical application of this philosophy. Built by our internal engineering team with direct support from Google’s AI engineers, it has become a powerhouse workflow tool. It handles the initial heavy lifting of summarisation, allowing our team to edit and rewrite with speed. Crucially, rather than replacing staff, the success of Bulletin has actually allowed us to create a number of new journalist roles – a fact that proves AI can be a net creator of value when deployed responsibly. When you close this divide, exceptional results follow.
Breaking the dependency on legacy software
One of the most liberating aspects of this tech transition is the potential to finally free ourselves from the industry-specific, multi-year contracts that have ultimately held us hostage for decades.
For too long, digital publishing has relied on massive, monolithic vendors for everything from content management systems (CMS) to digital asset management (DAM). We all know the drill. You sign a three-year deal, the implementation takes eighteen months, and by the time it is live, the technology is already becoming obsolete or your requirements have moved on. These vendors have largely sat on their hands, knowing the industry relies heavily on them and, most importantly, that migrating away is more than painful.
Their response to the AI boom has mostly been to slap a ‘Generate Text’ button on their existing legacy interfaces and charge a premium for it. But the ‘build vs buy’ equation has flipped. With the aid of AI coding assistants and open-weight models (like Llama or Mistral) that can be hosted privately, a small internal development team can now build bespoke tools in weeks that used to take vendors years to roadmap.
We are not quite at the point of ripping out the core ERP or HR systems. However, for editorial workflows, the era of the monolith is ending. We are moving toward ‘composable architectures’.
Instead of buying a bloated suite, we can build a specific ‘headline optimiser’ agent that runs on our own secure infrastructure. We can build a ‘tagging bot’ that learns our specific taxonomy better than any off-the-shelf software ever could.
We are nearing a future in which the ‘publisher’s stack’ is not a list of vendors but a library of proprietary agentic workflows. This allows us to own our destiny. If a new social platform emerges tomorrow, we don’t have to wait for our CMS vendor to build an integration. We simply build an agent to handle the formatting and API connection ourselves.
The CTO’s focus: productivity for the sake of journalism
As a CTO of a news media publisher, my obsession with agents, harnesses, and workflows is not remotely motivated by a desire to reduce headcount. It is born out of a desire to reduce drudgery.
If we’re honest, a significant portion of a modern journalist’s day is spent on tasks that require zero journalistic instinct. Formatting tables, transcribing interviews, tagging metadata, resizing images for five different aspect ratios, and rewriting the same blurb for three different newsletters. This is friction. It burns out talented people and distracts them from our company’s core proposition: finding and telling the truth.
My goal for 2026 is to deploy these agentic harnesses to handle the ‘commodity’ work.
If an agent can handle the initial pass of that internal revenue report, or if a harness can automatically transcribe and timestamp an interview, we give hours back to the human. That is time that can be spent picking up the phone, meeting a contact for coffee, or digging through a physical archive.
We use technology to buy time for humanity. That is the mission. We are building tools that allow our colleagues to stop acting like data processors and start acting like investigators and storytellers again.
The ‘discover apocalypse’ and the ‘direct relationship’
Finally, we must look at the elephant in the room for 2026: the stability of our traffic sources.
For years, many publishers have survived on the sugar rush of programmatic traffic, specifically from Google Discover. It has been a black box – a slot machine that occasionally pays out massive traffic for broad, sometimes low-quality content.
With the rollout of AI Overviews (formerly SGE) into search and the increasing ability of browsers to answer questions without a click, we must assume this ‘rented’ traffic will evaporate. Relying on an algorithm you cannot control to feed a business model you cannot predict is no longer a viable strategy.
The speculative horizon for 2026 suggests a ‘discover apocalypse’ where zero-click searches become the norm. The only defence against this is a pivot to direct, owned audiences.
Here, AI again can become the solution, not the problem. We can use the same agentic technology to hyper-personalise the experience for the users who do come to us directly.
Instead of sending one newsletter to 100,000 people, we should use agents to curate 100,000 individual newsletters based on each person’s reading history. Publishers who have already begun this shift, focusing on ‘habit’ metrics rather than ‘scale’ metrics, are finding that while their top-line traffic numbers might drop, their engagement and revenue per user stabilise.
Conclusion
The technology landscape of 2026 is not about finding magic AI software that does the work for us. It is about building a digital workforce that works with us.
By moving from prompts to harnesses, merging our skill trees, breaking our reliance on legacy vendors, and relentlessly focusing on removing drudgery, we can build media companies that are resilient, efficient, and, ironically, more human than they have been in years.
This article was first published in InPublishing magazine. If you would like to be added to the free mailing list to receive the magazine, please register here.
