Agentic harnesses and the future of DTP.
For two decades I have watched the DTP industry promise speed and quality through better software. The software arrived. The speed mostly did not. We have InDesign, FrameMaker, Acrobat Pro, XML toolchains, plug-ins for every conceivable check — and yet a typical regulatory submission still takes the same six to eight weeks it took when I started in 2006. The bottleneck has never really been the software. It has been the gap between what the software can do and what someone has to remember to make it do.
Agentic AI changes this. Not because the underlying model is dramatically smarter than the linters and macros we already had. It is, but that is not the point. The point is that an agent harness — a structure that lets the AI act, observe the result, decide what to do next, and keep going without being prompted at each step — closes the remembering-to-do-it gap. For the first time, the automation does not require a human to be in the loop on every step. It just requires the human to be in the loop on the decisions.
That is a much smaller, much more interesting role for the human. And it changes the economics of publishing operations in ways the industry has not yet fully absorbed.
What an agent harness actually is
Strip away the marketing language and an agent harness is three things wrapped around a language model: a set of tools it can call, a loop that lets it call them repeatedly, and a record of what happened. Give an agent a document and the right tools — read the file structure, query a style guide, check for accessibility issues, write a report, flag an exception — and it can work through a full pre-flight check without anyone telling it which tool to use next.
The autonomous part matters because publishing work is full of branches. A pharma patient information leaflet that fails an accessibility check needs different remediation depending on why it failed. A multilingual layout that overflows in German but not in French requires a different intervention than one that overflows in both. A traditional automation script would need explicit code for every branch. An agent reads the situation and chooses.
In our own work at B2K, we use IDML — the structured-text export format that InDesign produces — as the substrate the agent reads. IDML is XML. The agent reads it directly. It can identify every font in use, every text frame, every overflow risk, every paragraph style mismatch, every image without alt text. It does this in minutes, on files that would take a human operator three to four hours to inspect manually. And it never gets tired in the fourth hour of the afternoon.
The bottleneck has never really been the software. It has been the gap between what the software can do and what someone has to remember to make it do.
Why "autonomous" is the word that changes everything
The DTP industry has had automation for years. Preflight scripts in Acrobat. Find-change queries in InDesign. XSLT transformations on XML content. The reason none of this has compressed the production cycle is that all of it requires a human to initiate each step, interpret the result, and decide what to do next. The automation is a set of tools. The work is still done by the human moving between them.
An agentic harness inverts that. The human gives the agent a goal — "pre-flight this regulatory submission against the EMA template specification" — and the agent moves between the tools itself. It runs the font check. It reads the result. It identifies that Arial Hebrew is missing for the Hebrew variant of the document. It looks up the approved substitute in the style guide. It flags the substitution in the report. It moves on to the next check.
This is the production-floor reality that has changed in the last eighteen months and is changing more every quarter. The agent does not need to be brilliant. It needs to be diligent, and it needs to know which tools to call. The model already has the latter. The harness gives it the former.
What this means for a publishing operation
Three things shift, in order of how disruptive they are to current operating models.
1. The pre-flight stage compresses from days to minutes
Every regulated publishing workflow — pharma, journals, accessible book production — begins with a pre-flight check. Is the source file healthy enough to enter production? Are the fonts available? Does the layout conform to the template? Are the images at the right resolution and colour space? This stage used to occupy a senior operator for half a day per submission. An agent harness completes it in under five minutes and writes a structured report a project manager can act on immediately.
This is the single biggest unlock for a mid-market operation. It means the senior operator's time moves from checking files to making judgment calls on the files the agent has flagged. The operator does more, in less time, with less fatigue.
2. Multilingual rework collapses
The classic disaster of multilingual DTP is post-translation overflow: the layout was designed in English; the Arabic version is 30% longer; the German version is 25% longer; the Japanese version is 40% shorter; everything has to be reflowed by hand after translation comes back. An agent that reads the source IDML, calculates per-frame text capacity, and predicts overflow before translation starts can flag every at-risk frame with enough lead time for the layout team to design it differently.
I think of this as moving the problem upstream — from a rework cycle that happens after translation, to a design decision that happens before translation. It eliminates a category of work the industry has historically priced as part of the project. Clients notice quickly.
3. Accessibility compliance becomes routine instead of a project
The European Accessibility Act creates real compliance obligations for digital publications, and most publishing houses are not equipped to meet them at scale. The act requires every published digital document to meet specified accessibility standards — alt text on images, proper heading hierarchy, logical reading order, sufficient colour contrast. An agent can scan a publication against the standard, generate draft alt text for every image, identify reading-order issues, and produce a remediation plan in minutes. A human reviews and approves; the agent does the assembly. Compliance becomes a cost line, not a project.
Where humans still belong
I want to be specific about this, because the AI conversation has a tendency to swing between two extremes — either it changes nothing or it replaces everyone. Neither is right. Here is what I actually see, twenty years into doing this work and a year into running it with agentic tools.
The agent does the checking. The human makes the judgment. The agent reads ten thousand lines of IDML and flags forty issues. The human looks at the forty and decides which six are real, which thirty are noise, and which four are interesting enough to ask the client about. The agent never has the client context the human has built up over the engagement. The human never has the patience to read ten thousand lines of XML at three in the afternoon.
This is a healthier division of labour than the industry has had. The senior operator stops being a checking machine and starts being what the title actually implies — a senior practitioner who decides what good looks like. The agent absorbs the rote. Margins improve because the team is doing more meaningful work, not less.
What changes for clients
For the clients on the receiving end — pharma medical affairs teams, learning and development heads, production editors at journals — the visible changes are smaller than the operational ones. Turnaround compresses. A regulatory submission that took six weeks takes three. A multilingual eLearning rollout that took eight weeks takes five. The cost line gets smaller because the operator hours behind each deliverable are fewer. The quality report gets longer, because the agent leaves a full audit trail.
That last point matters more than it sounds. Pharma clients in particular have to demonstrate to regulators what checks were run on what files. A traditional production house produces a finished document and an invoice. An agentic operation produces a finished document, an invoice, and a structured log of every check that ran, every issue found, every decision made. The compliance value of that log is, in some cases, worth more than the production work itself.
The shape of the next five years
I think the publishing-services industry sorts itself into three groups over the next five years.
The largest agencies will build their own proprietary agentic platforms and price them as a premium feature. They will continue to win the biggest accounts because they have the brand and the procurement relationships. They will also continue to be slow, expensive, and structurally incentivised to keep human hours on the bill.
The freelancer end of the market will stay as it is — individual operators picking up small projects, working with their existing toolset, mostly not using agentic harnesses because they are designed for shop-floor workflows rather than solo work.
The middle — the mid-market publishing-services firms — is where the change will be sharpest. A firm of our size, with our depth of domain expertise and our ability to deploy agentic harnesses across the production stack, can deliver work at large-agency quality for mid-market pricing. The economics that supported the large agencies' premium pricing are eroding. The economics that constrained the mid-market firms' delivery quality are eroding too. The two are meeting in the middle, and there is a window — probably eighteen to thirty-six months — where firms that move quickly will define what the new mid-market looks like.
That is the bet B2K is making. It is a bet about division of labour, about who does the rote work and who makes the calls, about what publishing operations look like when the checking is autonomous and the judgment is concentrated. It is also, I think, the most interesting thing to be working on in our industry right now.
I'll write more about specific tools and workflows in the coming weeks — including how we use IDML as the substrate for the production agent, what an accessibility-checking agent actually looks like in practice, and where we think the limits of autonomous operation sit. If there's something specific you'd like me to write about, send me a line: prashanth.k@b2ktrans.com.