What is agentic media production?

By Ryan Hughes, Director of News & Sport Product Marketing

The word "agentic" gets used constantly and defined almost never.

I noticed it at an industry briefing last year. Several panels on agentic workflows, and nobody stopping to say what it meant in practice. I assumed that I'd just missed some quiet explainer, but I've since asked people across the industry, and almost all of them told me they'd never heard a clear definition either.

We used the word a lot at IBC2026 as part of our "agentic media production" theme, so it's worth being clear about what we mean by it - and why it matters for your production.

Let’s start with what “agentic” actually means

The word comes from “agency”—the capacity to act, to choose for yourself. So an AI agent works out its own steps. You give it a job, it decides which tools to use and in what order, and it adjusts as it goes. Traditional automation follows a route somebody mapped in advance. An agent finds the route.

Agency is the point. It's also the problem.

What that means for media production

Most AI in media production makes suggestions. It finds the clip. It drafts the summary. It tags the footage, transcribes the interview, recommends the shot. Then it hands the work back to a person.

That's been genuinely useful, and we've shipped plenty of it ourselves. Transcription alone has saved newsrooms thousands of hours. Object recognition has unearthed gems hidden within archives. AI helps, but a human still does most of the hard work.

Agentic media production is what happens when that changes. The AI agent takes a job and carries it through planning, preparing, checking, publishing, and telling you afterwards what the work cost. Under human control (ideally), but not waiting on a human at every step.

AI-assisted tools give you a recommendation. Agents give you a result.

What an AI agent actually does

An agent takes a job and works through the steps inside the systems you already use, stopping when a decision matters.

That might mean taking the day's footage, checking it against your technical spec, reconciling it with the existing project, and organizing it into bins with the metadata applied. It gives editors something ready to cut rather than something ready to sort.

It might mean searching your archive by what the content shows rather than what somebody named the file and screening out what doesn't answer the brief and telling you what it threw away and why.

Or it might mean reading the rights metadata before you publish, flagging third-party material inside a transcript, and reporting what's cleared and what's restricted.

Further upstream, you can see the same shape applying to planning and resourcing—turning incoming coverage into an assignment with people, kit, and cost attached before anyone's had to work it out by hand.

Each of those ends with something changed, with real work being done. That’s the difference.

Why this is harder for media production than most industries

Agentic AI is arriving everywhere, and most of it works on documents, tickets, and records. Media is a harder environment for three reasons.

• The actions are irreversible. Publishing to a platform, for example; you can delete, but you can’t reverse the fact that it was published. A misfire in a support workflow is embarrassing, but a misfire in a publishing workflow is a potential rights problem.
• The deadlines are real. A newsroom has a bulletin at six. A post house has a client waiting on a delivery. An agent that needs step-by-step supervision is slower than if you just did the job yourself, and that means it won’t get used.
• The judgment is craft. Which quote is aired. Where a cut lands. How a sequence builds. Those are decisions made by people with taste and experience, and automating them produces worse work.

The job for an agent is everything surrounding those decisions. The preparation, the checking, the versioning, the chasing, and the hours currently spent getting to the point where a decision can be made.

What agentic media production requires to be truly useful

Four things. Without them, you have automation with a chat window on the front.

Context. An agent is only useful with media it understands. Not just what a file is, but where it came from, what it was shot for, who approved it, what version it is, what rights are involved. Without that context, the agent is just guessing, and everything it produces would need to be checked by a human before anyone trusts it—putting you back where you started, plus an extra step.

Permissions. An agent should work within the boundaries of the person running it. If a journalist can't reach a confidential collection, neither should an agent acting on their behalf. That sounds obvious. It is frequently not how these systems are built, and you should ask.

Approval. Some actions can't be undone, and those need a human to confirm them. Because the agent is choosing its own steps, you can't approve the route in advance — so approval has to happen at the moment of action, enforced by the system rather than requested in an instruction. Instructions get misread, overridden, or lost halfway through a long conversation. A gate doesn't. If a vendor tells you their agent has been told not to do something, that isn't governance, it's hope.

Accountability. A record of what the agent did, what it read, what it changed, and who signed it off. In regulated production, that record is what an audit asks for. "We think it did roughly this" won't do.

Any vendor claiming agentic capability should be able to explain how they meet all four.

Where MCP servers fit

You'll have heard a great deal about the Model Context Protocol (MCP Servers), and rightly so.

MCP is a standard way for one AI application to discover and use another product's tools. It's open, quick to adopt, and lets an organization reach into several systems from one conversation. For research and individual productivity, it's often exactly right, and everyone building one is doing the industry a favor.

But it's a door, not a building.

An MCP server hands tools to an AI agent that lives somewhere else. The conversation, the reasoning, the memory, and the interface all stay outside your product. So does responsibility for supervision and history. The agent arrives each time without knowing what product it's in or what that product is for and works that out from scratch on every request.

This is fine for advisory work, but not so fine when the task will change production data.

MCP has a place, and most serious platforms will offer both. But with the ambiguity I mentioned at the start, it’s easy for the industry to blur the lines.

What AI agents still can't do

Agents are incredible, but not infallible.

They're bad at vague instructions. Ask an agent to "make me a social clip" and it will confidently offer you loudness targets, subtitle burn-in, and a color grade it has no ability to deliver. Ask it for in and out points and a caption, and you get exactly that. The distance between those two requests is currently the distance between useful and irritating.

The fix isn't a better model. It's refined, specific agents—a role with one job and a procedure to follow, rather than one generic assistant asked to do everything.

They don't know what they don't know. An agent that fails to find something will often tell you it doesn't exist. That's a claim about your entire archive based on one search, and in a newsroom, it has consequences. A journalist who’s told the footage isn't there stops looking for it.

They're only as good as the boundaries around them. An agent with no rules will do something nobody sanctioned, quickly and confidently. That’s why the four requirements above aren't a wish list.

Anyone selling you agentic AI who can't tell you where it falls short either hasn't built it or hasn't used it.

What separates good from adequate

Once the four requirements are met, four more things decide whether you're still happy in three years.

You keep what it learns. Correct an agent once and it should remember. That accumulated knowledge about how your newsroom actually works is valuable. It should belong to you, not to whichever model provider you happened to pick.

You choose the model. Better models arrive constantly. Adopting one shouldn't mean rebuilding workflows you've spent a year refining.

You can see the cost. AI consumes compute, and compute costs money. At production volumes it adds up fast. You should see what agents are spending before the invoice arrives, not after.

It works where your media is. A broadcaster with petabytes on-premises can't easily upload them to make AI possible. The processing must come to the content.

What to ask any vendor claiming agentic AI

• Where does the reasoning happen? Inside the product or in a chat window connected to it?
• What context does the agent start with? Does it know what it's working on or does it have to be told every time?
• Whose permissions does it use? And can you tell afterwards whether a person or an agent made a change?
• What stops it doing something irreversible? Is that enforced by the system or written into an instruction?
• What does it cost to run? Can you see that before you spend it?
• What happens if you change AI provider? Does everything you've built come with you?
• And one more: Where does it fall short? Anyone who says “nowhere” is either not listening or not telling you.

None of those require a technical background to ask. If a vendor can't answer them in plain language, that tells you something, too.

Agentic media production is going to change how this industry works. It's worth being specific about what we mean by it before everyone claims to have it.

I'd have found that useful, sitting in that room last year.