The Orchestrator and the Specialist: Why the Toolkit Didn't Shrink

The Orchestrator–Specialist Split is the market structure in which generic AI models act as planners and coordinators while specialised tools compete to remain the callable expert within their own domain. The arrival of capable generic AI models was supposed to shrink the toolkit. One model, the argument ran, would absorb the functions of the specialised software around it — the design tool, the analytics platform, the project tracker — until the everyday stack collapsed into a single interface. The opposite is happening. Specialised tools are not disappearing; they are repositioning. They concede the planning layer to the generic model and compete to remain the expert it calls. The toolkit is not shrinking. It is reorganising.

The consolidation thesis deserves its strongest form. Generic models improve on a steep curve; they already write, analyse, and code across domains that once required separate products. If capability were the only variable, consolidation would follow — the general model would close each gap in turn and the specialised vendor would become a feature. But capability inside a domain compounds differently than capability across domains. The specialised vendor holds proprietary data, embedded workflows, domain-specific evaluation, and regulatory depth that a generic model reconstructs slowly, if at all. The vendor's rational strategy is not to defend the interface. It is to defend the gap.

Two motives, one behaviour

The repositioning runs on two motives that produce identical behaviour. The defensive motive is survival: a tool a generic model cannot call is a tool the generic model eventually replaces. The offensive motive is position: the tool an orchestrator calls by preference becomes the expert endpoint in every workflow the orchestrator plans. Under both motives, the specialised vendor invests in the same two assets — an agentic interface, and the specialist gap, the capability delta the tool maintains over a generic model within its own domain.

The structural evidence is the speed of interface adoption. Within months of agent-facing interface standards emerging, specialised vendors across design, project management, analytics, and documentation shipped machine-callable endpoints — not because their human users demanded them, but because absence from the orchestrator's reach is absence from the workflow. This is the same architecture specified for use inside an individual business's own stack as Agent Specialisation — designing agents with a narrowly defined task domain, a constrained capability set, and a formal cooperation interface, so each agent outperforms a general-purpose one within its domain while remaining composable with the rest. The Orchestrator–Specialist Split is that same architecture playing out at market scale rather than inside a single company's stack: generic models hold the coordination role, specialised tools hold the execution domains, and the interface between them is the competitive surface. Both instances map onto the same underlying Task Tiers (T1 / T2 / T3) structure — the orchestrator coordinating across tiers, the specialist executing at the task level where domain depth actually determines competitive position.

The pricing consequence

A market that reorganises its division of labour reorganises its pricing. The seat licence prices software against human headcount — a unit of consumption that loses its meaning when agents perform the work, the demand-side argument already made in The Death of the Seat License. The market's default successor narrative is outcome-based pricing: bill for results, not access. Outcome-based pricing inherits an attribution problem it cannot solve — proving which tool in a multi-tool workflow produced which result.

Agentic Consumption Pricing resolves what outcome-based pricing cannot: cost scales with agent call volume and the capability tiers exposed to agents, shifting the unit of consumption from human seat to agent transaction. This is the vendor-side pricing response to the same shift buyers are practising as De-SaaS-ing — the seat-based cost structure named directly by the UI Tax is what Agentic Consumption Pricing replaces. It is operationalised through the Machine-Readable Interface (MRI): the same structured, API-first layer that lets an external agent discover and transact with the tool is what generates the transaction volume Agentic Consumption Pricing meters, requiring no attribution because the unit being billed is the machine's own consumption. Vendors that reprice this way retain the customer at the interface layer rather than losing them at the application layer.

Who operates the toolkit

The reorganisation changes the operator, not just the stack. In the split, the human does not use twenty tools; the human governs the orchestrator that calls them. What varies is supervision — how much guidance the plan requires, how much final judgement the output demands. The Stewardship Model is the Arco operating model in which a single competent operator oversees an agentic stack, acting as architect and exception handler rather than executor — and this is that model in its next environment, governing an orchestrating model that routes execution to specialised endpoints rather than executing the work directly. A stack built correctly on the split is measured the same way every Stewardship Model deployment is measured: by MTTI (Mean Time to Intervention) — how rarely the system needs the Steward at all, with the target set above 72 hours on core workflows.

The Operator's Verdict

We do not ask which tools survive consolidation, because consolidation is not what is happening. We ask which layer of the split each business owns — the plan, the domain, or the interface between them. The specialised vendors understood this before their investors did. Generic AI competes for the plan. Specialised AI competes for the call. The margin lives at the interface where the two meet.

Technology changes who can build the plan. Architecture determines who still gets called to execute it.