Agentic Workflows in Media, and What They Ask of the Network

An agentic workflow is software given a goal, such as finding every take of a line in a series’ rushes or conforming and delivering tonight’s cut, that works through the steps itself.

In media that puts three demands on the infrastructure underneath: storage that takes bursts of reads and writes, access permissions scoped to a job rather than a person, and a fast, predictable path between the files and the processing. Two of the ten principles in MovieLabs’ 2030 Vision, published in 2019, already point at the last two.

‍

Agent language has moved into the product pages themselves

Looking across one product page each from thirty-four technology companies, most of them selling storage, file transfer, asset management, review tools or networks to media businesses, between mid-July and mid-September 2026, by our count, eighteen of those pages changed in ways that said something about how the company wants to be seen.

Seven of the eighteen gave AI more room on the page. Eight reorganised the page, most of them around workflows, use cases and industries instead of features, so a visitor now starts from a job to be done rather than a feature list.

The word “agentic” itself did not spread far: on 5 October 2026 four of the thirty-four pages used it, about the same number as in August. What changed is where it sits. Over the summer it moved from banners and featured content into the structure of the pages, as a new menu item for agentic workflows on one page, and as the headline of a product announcement on another.

IBC, held in Amsterdam from 11 to 14 September 2026, sat in the middle of this, and eight of the eighteen pages put up a banner for it. By 5 October, five of those eight no longer mentioned the show at all.

The agent language stayed, and some was added after it: one page renamed its “Key features” section “Our agents” within two days of the show closing. On 5 October 2026, five of the thirty-four pages talked about AI agents in their navigation, headings or featured content.

None of the five pages that talked about AI agents on 5 October 2026 mentioned bandwidth, latency or throughput. They describe what an agent will do for a team. They say little about what it will ask of the storage, the permissions and the network underneath it.

A product page is one surface, and a menu label is not a shipped capability, so this is a reading of what companies chose to say, not of what their software does. What agentic post-production means for the people in a facility has been covered before. This article is about the other half.

‍

The trade press asked for a reality check

IBC’s own editorial site took a different line going into the show. On 25 August 2026 IBC365 published a feature headed “IBC Content Everywhere: AI gets a reality check”. On 28 September 2026, after the show, IBC365 followed it with “Content Everywhere at IBC2026: AI moves from hype to action”.

Both sit behind registration, so the headlines are what can be quoted here. Read as headlines, they describe a movement: from scepticism before the show to action after it.

Both sides are right about something. The vendors are right that post-production buys a way of getting a job done, not a list of features. The press is right that a label on a menu proves nothing.

And “action” is where the infrastructure question starts, because an agent that acts on media files is a new kind of workload with a shape of its own. It can only be as reliable as the storage it reads, the permissions it holds and the path its files travel.

‍

Storage: an agent reads in bursts, not at a human pace

An editor working on a project reads files at the pace of a person: a clip, a scrub, a pause, a decision. An agent asked to transcribe a week of dailies, or to make proxies of a day’s camera originals, reads all of it as fast as the storage will give it up, and then writes its results back.

Three things follow from that. Load arrives in bursts, and often outside working hours when jobs are scheduled, so storage sized for a day of editorial work can become the bottleneck at night.

Versions multiply, because every pass writes something new: a transcode, a proxy, a conform, a quality-control report, each needing a name, a place and a record of what produced it.

And unless something is watching for it, an agent that cannot reach its files does not phone anybody. It times out, retries or finishes late, and the delay surfaces further down the schedule.

‍

Access permissions: scoped to the job, not the person

Most permission models in post-production are built around people: an editor, an assistant, a vendor’s team, each given access to a project.

MovieLabs, a non-profit technology lab run jointly by major Hollywood studios, made this one of the ten principles of its 2030 Vision in 2019: every individual on a project is identified and verified, and their access permissions are efficiently and consistently managed.

MovieLabs’ later work on production security proposed that production move to a zero trust security architecture, the model NIST set out in Special Publication 800-207 in August 2020, in which no asset or account is trusted because of where it sits on a network.

The principle is written about people. Applied to an agent, it still reads well. An agent needs its own identity rather than a person’s login. It needs access to the files its job needs, for as long as the job runs, and no longer. And it needs a record of what it opened, changed and sent.

An agent running on a shared account with a person’s permissions is the opposite of all three: it can reach everything that person can, and the audit trail cannot tell the two apart. The same thinking sits behind zero trust in post-production generally.

‍

The network: the path between the files and the processing

Another of the ten principles in MovieLabs’ 2030 Vision is that applications come to the media: the processing moves to where the files are, rather than the files moving to the processing. Where that holds, an agent’s reads and writes stay close to the storage.

Where it does not, because the files sit in a facility and the agent runs in a cloud region, or the other way round, every task the agent performs becomes a transfer, and the connection between the two sets the pace of the whole workflow.

For an agent, a predictable path can matter as much as a fast one. A person notices when a transfer is slow and changes the plan. A scheduled job assumes its inputs arrive when they always have.

If the path between storage and compute is shared with the working day’s traffic, or crosses the public internet, its timing varies, and every step downstream inherits the variation.

‍

Before an agent touches production files

The questions worth asking of any agentic tool are as much about infrastructure as about AI:

1. Where will it run, and where are the files? If the answer is two different places, the connection between them is part of the product, whoever supplies it.
2. What does the storage do when the agent and the editors want it at the same time? Burst load from a job should not slow the people working on the project.
3. Whose permissions does it use? An agent should have its own identity, access scoped to its job, and a record of what it touched.
4. What happens when it cannot reach its files? A tool that fails quietly needs someone, or something, watching for it.

‍

Where Sohonet sits

Sohonet has supplied network, cloud access and data-movement infrastructure to film, television and post-production for 30 years, which puts it on the infrastructure side of this question rather than the AI side. What it sees in its customers matches the pattern above.

Sohonet’s chief executive, Chuck Parker, described it in January 2026: customers keep latency-sensitive editorial, colour and VFX work on premises, where performance and cost are predictable, and use the cloud for elasticity and for bringing in new vendors and teams quickly. What they ask for is a connectivity and data-movement layer that lets them shift work between the two without re-architecting every show.

The Sohonet Media Fabric network, with its direct cloud access, and Core, Sohonet’s production asset management system for storing and finding a production’s files, are the layers this article has been about. Sohonet joined the Leadership Council of the MovieLabs Industry Forum on 5 August 2026.

‍