Verdict
September’s evidence suggests that an agent is no longer just a product attached to a model. It is a route through model access, data, an end-user surface, connected applications, and the right to take action. That route is increasingly shaped by commercial terms, marketplace rules, managed surfaces, and approval design.
The claim is not that platform lock-in or market concentration is already proven. It is that builders now need to treat the route as an operating dependency. An independent builder’s defensible asset is a portable workflow package: task contract, durable state, policy, acceptance checks, and a tested right to replace the route.
Central contradiction
More useful agent reach requires less unilateral control by the builder over where the agent acts, under whose terms, and with which permissions.
Observed, inferred, forecast
Observed. Model access inside a product can be contingent on a commercial and policy decision: OpenAI said it intended to wind down Cursor’s contract after its change of control (OpenAI). Claude’s marketplace and plugin-submission path make discovery and listing a distinct layer (marketplace, plugin documentation). Product descriptions for Dots and Muse place connected data, review, interruption, or approval alongside action. A qualitative study of persistent team agents also found that acceptance depended on relational boundaries and gradually earned authority, not task completion alone (paper).
Inferred. The deployable product unit is a *governed route*: capability, identity, data scope, distribution surface, policy, and consequence connected in one path. API-key portability is not workflow portability; a move must preserve state, policy, observability, approvals, and action semantics as well as prompts.
Forecast. Over the next year, prominent agent platforms are likely to document more explicit eligibility, review, distribution, and per-action authority rules. The counterweight will be portable workflow packages that preserve more than a prompt and a key. This is an early forecast, not evidence of adoption.
Triangle
| Layer | Reading | | --- | --- | | Foundation | Agents can be delivered through managed marketplaces, enterprise homes, connected-app surfaces, and cloud computers; a workflow may have more than one model-access path. | | Dynamics | Builders must design for portability and staged delegation, while teams encounter platform policy at distribution and action time. | | Emergence | A governed route becomes the unit of deployment: a permitted, observable, replaceable path from context to action. |
Nine cells
| Cell | Claim | Status | | --- | --- | --- | | A1 | Model access is a commercial and policy dependency, not a permanent component property. | Observed | | A2 | Marketplaces and managed homes can combine discovery, routing, policy, and administration. | Observed | | A3 | A workflow can have plural access paths, though semantic portability remains incomplete. | Observed, limited | | B1 | Builders must model provider, surface, connector, eligibility, and action right alongside agent logic. | Inferred | | B2 | Consequential authority is staged as prepare, review, approve, then execute. | Observed | | B3 | Buyers will value preserved discretion and an exit path, not capability alone. | Inferred, medium confidence | | C1 | The deployable unit becomes a governed route joining capability, identity, data scope, surface, and permission. | Inferred, strong | | C2 | Distribution control becomes a first-order product boundary for agent builders. | Inferred, medium-high confidence | | C3 | Portable packages—state, policy, capability contract, and test fixtures—become a counterweight to route concentration. | Forecast, early |
Edges
- A1 → B1 — strong. Conditional model access turns provider terms into an architecture input.
- A2 → B1 — strong. When discovery, routing, and administration share a surface, integration becomes distribution.
- A3 → B3 — medium. A second access path makes an exit test possible, but does not prove semantic portability.
- B1 → C1 — strong. Mapping the dependencies makes the route the object a team deploys and governs.
- B2 → C1 — medium-high. Staged authority makes permission and consequence part of the route.
- B3 → C3 — medium. Demand for an exit path creates pressure to package state and policy independently.
- C2 → A2 — reinforcing. Strategic distribution encourages more catalogue, routing, and governance features.
- C3 → A3 — forecast. Mature packages would retain more across a move than prompts and an API key.
Counter-evidence
Plural access is a real counter-branch. Kilo’s portability argument explicitly frames provider choice as resilience (Kilo), and a model can appear across several surfaces (Anthropic’s Fable and Mythos documentation). Neither is proof that production state, policy, approvals, and audit records move intact.
The marketplace and managed-surface evidence is largely launch or product-description evidence. It supplies no supplied measure of migration time, churn, listing rejection, reach, connected-action error, approval burden, or comparative adoption. The team-agent study is one qualitative study, not a distribution study. The correct reading is a design and operating-model signal, not a demonstrated market structure.
Rejected hypotheses
- “The best model automatically owns distribution.” Rejected: access can be contractual, while discovery and routing sit in separate layers.
- “A marketplace listing proves lock-in.” Rejected: listing alone says nothing about ranking, fees, switching cost, or reach.
- “A permission dialog solves trust.” Rejected: authority also depends on tacit norms, relational boundaries, and the quality of review.
- “API-key portability equals workflow portability.” Rejected: policy, state, credentials boundaries, and action semantics must move too.
Revision conditions
Strengthen this mechanism if three independent builders publicly move the same agent workflow between platforms while retaining state, policy, auditability, and action semantics, or if marketplace rule changes demonstrably alter an established agent’s reach.
Weaken it if teams routinely move production agents between providers and surfaces through standard interfaces without material loss of connectors, policy, state, or distribution.
Falsify the concentration branch if connected agents gain wide use while access, routing, and action rights remain materially independent of platform terms, channels, and policy layers.
Builder takeaway
- Inventory one live workflow’s route: model provider, surface, identity, connectors, data scope, listing terms, approval owner, and write rights.
- Extract a portable capability package: task contract, prompts or skills, durable-state schema, policy rules, fixtures, and acceptance checks. Keep credentials and adapters outside it.
- For one consequential action, implement draft → explicit review → execute and record who grants each transition.
- Run a non-production exit rehearsal: replace one model or surface and record every state, permission, audit record, or UX element that cannot move. That gap is the portability backlog.