Editors’ note: Today’s post is by Chef Hong Zhou and guest blogger Adam Hyde, founder of Pure Science Inc.

If an agent can generate software on demand, why rent a fixed product from a vendor? The question has a name — the “SaaSpocalypse” — and to many it looks like an existential threat to every large platform that bundles interface, logic, data, and workflow into one product sold by the seat. Scholarly publishing is built almost entirely on platforms of this kind.

The PurePub.ai event (held on May 22, 2026) took it head-on — with Adam Hyde moderating a discussion with Hong Zhou of KnowledgeWorks Global, Stuart Leitch of Silverchair, and Paul Shannon of eLife. The conversation ranged widely, but one of Hyde’s questions was especially direct: Are agents “a full frontal attack on SaaS?”

The panel’s conclusion was more interesting than a yes-or-no answer; it is very likely that the platform isn’t dying, but is being split in two. What was a single bundle comes apart into a trusted core below and an agentic layer above — and the seat and the screen that used to join them lose their hold.

The Layers Separate

Although we all agree the future is unknowable, Leitch laid out one trajectory directly. Platforms could be “broken down into primitives” — at the base, “databases with the business rules … enforced within them;” above them, “progressively this kind of agentic layer,” with ownership of that layer still unsettled.

They separate because the consumer changes. For three decades the consumer of a platform was a person sitting at a screen (a seat). Agents alter that scenario in stages: first humans use the system directly (what we have to date); then agents appear as consumers alongside them, then, whether the human or the agent is doing the work, the data passes through agents that communicate directly with the headless core rather than through the legacy interface.

The core, as they say,  goes “headless” (a technical term meaning there is no primary user interface tied to the core), according to Leitch — not that the screen vanishes; the agent works exclusively through the core to accomplish its objective. The user interface, no longer the only way in, detaches from the platform: free to live anywhere, assembled by anyone. Shannon drew the first-stage consequence for the people still in the loop: we will need “a different architecture for the humans to interact with” with agents first, feeding data to humans.

What the Interface Was Carrying

That detachment takes something less obvious with it. The interface was holding two things together: trust, and the way the whole arrangement was paid for.

Trust in the interface is due partly, as Hyde put it, to the fact that “you look at an interface, and it’s the same interface you looked at yesterday.” You see the data, you know how it behaves, and that confidence was built slowly, over years. If the interface is no longer bound to the core, it can be regenerated, swapped, or rebuilt on demand, and that familiarity no longer vouches for what sits beneath it.

The second is more serious: the business model. The seat counted users, and that worked while every user was a person sitting at a screen (a ‘seat’) — in this (current) scenario, headcount and cost rose together. Detach the screen, put an agent in front of the core, and the two come apart: a few people directing agents can drive load no ‘seat count’ would be able to adequately describe.

“I don’t think it’s a seat model,” Hyde said. So, pricing moves from subscription toward consumption, and from consumption toward value. For Leitch, the unit is compute, not headcount, that wins. The seat simply measures the wrong thing. SaaS platforms, in his view, shift from a per-seat model to some form of consumption model. Perhaps with a BYOK model (‘bring your own key,’ where customers supply their own model API keys).

There are other scenarios. Charging per token, or charging for outcomes, is the direction some AI software is taking. With open-source models, some (e.g., OpenCode, the open-source coding agent tool) are making good margins on tokens, while still providing near-frontier outcomes and being significantly cheaper. The compute pricing model (Palantir is particularly based on this) is also interesting and has some advantages: it’s legible to procurement, and it sits on a known, understood infrastructure-pricing model (tiered pricing for infrastructure) — keeping the business out of the token-markup game, or “charging for outcomes,” which looks to be a volatile place to be.

Which Layer Resolves First?

If price follows value, the question perhaps becomes where value now sits? One argument is that it sits in the trusted core, and the agentic layer (with whatever holds the trusted record at the base) and the agents above the core. The agent (or “orchestration”) layer is, in Leitch’s words, “up for grabs.”

The two layers do not resolve on the same timetable. The base can now be logically analyzed, as the forces acting on it are already visible (agents are increasingly becoming the consumers). The contest for the layer above is still open and turns on a question this post sets up, but leaves to a later article.

The Trusted Core

In the panel, we also turned to the legacy incumbents. The challenge is that the factors holding customers in place are weakening — the cost of switching platforms is falling on two fronts, making it easier for an organization to walk away.

The first is the cost of building your own. Standing up a submission system, hosting platform, or peer-review workflow has traditionally required significant engineering investment. The impact of AI-assisted development on software productivity remains contested. Studies conducted through 2025 reported everything from substantial gains in coding speed and task completion to slowdowns among experienced developers working in mature codebases, where review and verification overhead could outweigh productivity benefits.

At the same time, many of the most-cited studies evaluated earlier generations of AI coding tools. Since late 2025, advances in AI coding agents and agentic development workflows have expanded the scope from code completion toward systems capable of operating across entire repositories, executing tools, running tests, and iterating on solutions

While rigorous evidence is still catching up, there is growing reason to believe that development productivity is increasing, and the cost of creating software is falling, particularly for new applications and routine development tasks. More importantly, industry participants are increasingly behaving as if that is true. For incumbent providers, this presents both a threat and an opportunity.

As Leitch put it, his advantage is that he builds from “the full codebase” and the “history of the evolution of the codebase,” allowing him to refactor faster and draw on decades of accumulated business logic. Which is why his instinct is to move first: if software can be built more cheaply and rapidly, he intends to “disrupt ourselves at that same speed” before someone else does.

The second front is migration. A system exposed for agents in an ‘agent first’ user environment is, by that very fact, a system agents can read, extract, and map onto other data models. Migration cost, Leitch expected, will “come down quite a lot.”

Two challenges for publishers that form part of the SaaS moat are diminishing. What remains is the system as the proven record — the thing an institution has run its operation on, and audited, for years, plus the relationships, industry knowledge, and support services that the incumbent carries forward. “The trust aspect,” Leitch said, is likely to be something that ”retains essential value in the core system” (quote modified slightly by Leitch after the discussion). And scholarly publishing rewards it directly: the field is “a bulwark against change” (although conversations in general at the PurePub.ai conference were quite forward-looking and insightful in anticipating the coming changes caused by AI).

Where Platform Value Moves Next

This returns us to the question: are agents a full-frontal attack on SaaS? And, by extension, what should publishers and platform vendors do next?

From the PurePub.ai discussion, the consensus was no — the attack is narrower than that. At the infrastructure layer, the SaaS platform endures as the trusted system of record, along with the relationships, integrations, services, and institutional trust that have accumulated around it. What is being challenged is not the existence of the platform itself, but the assumption that users must interact with it through a vendor-controlled interface and workflow.

If scholarly platforms are splitting into a trusted core and an agentic layer, the strategic question for publishers is not whether one replaces the other, but how value is distributed between them. The record, workflows, and institutional knowledge that underpin scholarly communication may remain highly valuable. But if agents increasingly mediate not only how research is discovered and consumed, but also how it is reviewed, assessed, produced, and managed, then a growing share of value may accrue to the agentic layer. Understanding where value resides — and how the two layers reinforce each other — may become one of the central strategic challenges for publishers.

In part two of this post, we will turn to the agentic layer itself and examine what happens when the interface becomes intelligent, personalized, and increasingly independent of the platform beneath it.

Authors’ note: AI was used to polish and critique this essay. All the information is verified by the authors.