AI agents are becoming the new browser
The French competition authority published an opinion on the AI agent market yesterday. The number that jumps off the page is concentration: OpenAI, Google, and Anthropic account for more than 84% of users, according to the authority.
That is striking for a category that still feels like it was invented last Tuesday. It also frames agents more usefully than another capability benchmark does: as a new gateway to the digital economy.
The browser took you to services. The agent increasingly chooses and operates them for you.
The interface becomes the market
Ask a browser for running shoes and it gives you pages, ads, shops, and space to make a bad decision yourself. Ask an agent and it may compare products, apply preferences, select a seller, and eventually complete the purchase without exposing most of the market it considered.
That is wonderful UX. It is also an enormous allocation of power.
Which merchants get a tool integration? Which result is described first? Does the agent prefer a service owned by its platform? Can a smaller provider tell whether it was excluded for quality, price, latency, or a private commercial agreement? Once the agent acts, recommendation and execution collapse into the same invisible step.
Search had versions of these problems. Agentic commerce adds write access.
Open protocols help, but governance matters
MCP and similar protocols lower the technical cost of connecting a service to an agent. That is necessary and not sufficient. A protocol can be open while discovery is closed, ranking is opaque, and the dominant client controls which capabilities users ever see.
USB-C did not decide which device you were allowed to plug in. An agent platform can.
The competition authority also points at interoperability and the governance of technical standards. This is where the boring details become constitutional: who can propose changes, whether multiple clients implement the same semantics, how identity and permissions move between them, and whether a user can export the context that makes one agent useful.
A standard owned in practice by one distribution channel is an integration API, not an open ecosystem.
What builders should keep portable
For teams building services that agents can use, I would avoid making the agent platform the system of record.
- Keep authorization and consent in your own service.
- Make tools deterministic enough to work across clients.
- Return prices, provenance, and constraints as structured data rather than persuasive prose.
- Log what the agent requested and what your service actually did.
- Support revocation and idempotency as first-class operations.
Most importantly, preserve a direct path for the human. An agent should be a powerful client of your product, not the owner of the customer relationship.
The agent wars look like a model race today. If agents become how people reach software, information, and commerce, the real contest is browser-sized — and we should treat it that seriously before the defaults harden.