The MCP Eliminates Its Sessions, Not Business Requirements

Recent modifications to the Model Context Protocol (MCP) remove sessions from the protocol to facilitate the distribution of requests among servers. This protocol connects AI applications to enterprise services. Its stateless operation does not exempt organizations from retaining business data, managing interruptions, and controlling access.
A command can be executed twice
An ERP may have created an order while the agent never received its response. A connection drop is enough to create this uncertainty: retrying the call risks duplicating the purchase. If another MCP instance takes over after a failure, it must be able to establish what has already been executed. Therefore, improved routing availability does not guarantee the successful completion of business processing.
Idempotence serves as a possible safeguard: a unique identifier for the operation is associated with a saved result. Upon a new submission, the application can return this result without re-executing the action. In the absence of a retained result, the service must query the ERP before any retry.
Long-running processes can also remain partially executed, for example, when stock has been reserved without the order being created. Resuming them requires a history of completed steps, typically maintained by business applications. Multi Round-Trip Requests support exchanges requiring additional information; the optional Tasks function ensures tracking of long operations.
Monitoring must follow business results
A gateway, an MCP server, and the business system can all report operational status while the transaction passing through them fails. To diagnose this case, distributed tracing links events that occurred across different services. The MCP can carry standard trace information, usable to track the journey from the agent to the business system. However, this tracing must be configured in each of the relevant systems.
Dashboards should separate the success of a protocol response from the actual completion of a business operation. Tool delays, denied authentications, new attempts, and slowness of the business system can all contribute to diagnostics. For sensitive actions, audit trails should specify the requester, the affected resource, and the result.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Requests can change servers
Previously, MCP servers generally provided an identifier to recognize a session during subsequent calls. If its state was retained in memory, those calls had to be directed to the same instance or share that state among servers. A failure could interrupt the exchange. Now, the information necessary for processing accompanies each request, which can be received by another instance.
This autonomy facilitates capacity addition, replacement of a failing instance, and the use of serverless infrastructures. It does not eliminate the memory required for the business process: a tool can, for example, save a draft of a purchase order in the ERP and then return a reference. Subsequent calls can use this reference to consult or modify the draft, even from another MCP server. Shared storage can also make records accessible to different instances.
Before changing infrastructure, organizations should verify the compatibility of legacy clients, especially those that depend on sessions. Preparing for a migration should also identify the location of business data, test retry mechanisms, and ensure that each transaction remains tracked.
Permissions remain controlled in applications
For its protected HTTP connections, the MCP relies on OAuth: each call carries an access token. The server validates it to verify the identity of the caller, the recipient service, and the permissions granted. Multiple agents can use a common MCP environment, but the indications a client provides about its software do not prove its identity or rights. Consulting or modifying a customer record always requires checking the permissions of the user or the workload that initiated the request.
A gateway can impose common rules upstream of the servers. Headers identifying the operation allow for directing calls and limiting sensitive or heavily requested functions without examining the full content of the request. These checks do not replace those of the business software: rights must be applied consistently at the entry point as well as within the application.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.