Marketing got its app-store moment this month, twice. In the first days of July, TikTok opened an Agentic Hub, a marketplace of AI skills for advertisers. Three weeks later, HubSpot shipped Agent Hub, a home for every agent a go-to-market team runs, with a marketplace inside.
Our argument: the strategic story in both launches is distribution economics, not features. The lock-in is not a side effect the platforms will get around to fixing. It is the point. The app-store era already taught this lesson to anyone who sold software through one, and it is worth replaying before your agents move into someone's hub.
What launched, specifically
TikTok's Agentic Hub is a marketplace of native and third-party AI skills for campaign creation, creative, catalog management, and performance analysis, wired to TikTok's ad platform through its Model Context Protocol server. Partners with published skills at launch included HubSpot, Wix, and Constant Contact.
HubSpot's Agent Hub, in public beta for Professional and Enterprise customers, is pitched as one home to see and run every agent across the business, with Agent Builder and an Agent Marketplace inside it. The load-bearing idea in the announcement is context: agents built there arrive already wired to deal history, contact records, call records, and buying signals: no separate setup, no field mapping. Custom agents meter against HubSpot Credits when they act.
Note the cross-listing: HubSpot operates a hub and is simultaneously a skill inside TikTok's. Every platform wants to be the store; every platform is also an app in the other store. That tells you how much everyone believes distribution is the prize.
Disclosure: this publication is produced with support from Docket, which sells software in this category, including agents that live outside anyone's hub. Weigh the argument knowing where we sit.
The pattern is older than the products
App stores ran the same sequence: distribution first, rules later. In 2011, Apple required apps selling digital goods to use its in-app purchase system, with its 30% commission, or strip out links to buy anywhere else. Amazon responded by removing the Kindle app's store button, and buying a book inside that app quietly stopped being possible. It stayed impossible until 2025, when a federal court ruling finally let a purchase button back in.
Fourteen years. That is how long the rent lasted once the dependence was set, and it was set back when the store was young, generous, and mostly about discovery. The rules arrived after the audience did.
Where the lock-in actually lives
The app-store analogy undersells one thing. An app was a binary you installed; leaving a store meant losing distribution. An agent is a service that accretes operating context, and that context is the platform's real moat, not the listing.
Read HubSpot's pitch again: the reason to build in their hub is that agents arrive already knowing your deal history, contacts, calls, and buying signals. That is genuine convenience, and our view is that it compounds into the exit cost. Every day an agent runs, it deepens its dependence on context that lives in the hub: its memory, its workflows, its performance history, its metering. Leaving does not mean reinstalling somewhere else. It means re-integrating everything the agent knows, while the meter that prices each action belongs to the landlord.
What the hubs genuinely fix
The hubs are answering a real problem. Agent sprawl is not invented: teams are running agents from many vendors with no shared view of what is live, what it did, and what it cost. A console with status, outcomes, and one-click activation is real operational value, and so is marketplace vetting. For a lean ops team, hub-first is a rational way to start.
The analogy also cuts against itself in one place: TikTok built its hub on the Model Context Protocol, an open standard. A skill published over a protocol is more portable than an app-era binary ever was, since capability can move between surfaces. What the protocol does not port is accumulated context. The skill travels; what it learned about your business stays home.
Six questions before you commit
Ask these of any hub. The vendor's answers are the exit map, and none of them get easier to ask after a year of usage. They are ownership questions in disguise: a hub that answers them for you has taken the operating decisions along with them.
Where does the agent's memory live, and what does an export contain: configuration, learned context, conversation history, or just settings?
Whose records does it write to? Your CRM objects, or the hub's own objects that reference yours?
Whose identity does it act under? An identity you govern, or a platform service account you rent?
What does the same action cost outside the hub? Metered credits are a take rate on execution; price the alternative before the meter becomes the budget baseline.
Which workflows die if you leave? Count the automations that reference hub-native objects. That number is your real switching cost, in rebuild-weeks.
Does the vendor exist outside the hub? An agent that is only distributed inside one platform inherits that platform's rules, pricing, and lifespan. See the Kindle button.
FAQ
Should we avoid agent hubs? No. Enter them the way experienced developers entered app stores: for the distribution and the console, with an exit map written before the dependence forms.
Is MCP the escape hatch? Partially. Protocols make capability portable across surfaces. They do not make context portable. The agent's accumulated knowledge of your business stays wherever it was allowed to accrete. Insist on knowing where that is.
Native platform agents or independent vendors? Match depth to stakes. Platform-native agents are strongest at convenience tasks inside their own data. Specialized independents earn their keep where the work is high-stakes and the knowledge required is deep, and where you want the agent's memory to be yours.
How is this different from the app-store era? It is the same playbook with heavier gravity. Apps held code; agents hold accumulating operational context. The store owned your distribution. The hub, left unexamined, owns your agent's memory. Which makes the real design question not whether to enter one, but which parts of the loop you keep.