Autonomous software agents—programs performing tasks without direct human oversight—hit a digital wall when they reach for protected resources. The wall is the standard login flow: the sequence of steps that verifies who is asking and decides what they may do. Built for a person at a keyboard, it assumes someone is there to click a button.
The Model Context Protocol (MCP), a framework for connecting AI assistants to data sources, shows the friction plainly. The project’s own OAuth Client Credentials extension page states that the standard flow requires a user to interactively approve access: “a browser opens, the user logs in, and grants permission … it breaks down when there’s no user present.”
So the workarounds start. The MCP authorization specification makes authorization optional; systems on STDIO, or standard input/output—a direct channel between programs—are told to pull credentials straight from the environment instead. Nothing authenticates. A key was simply left on the desk.
The second escape is the client credentials extension: opt-in and, in the project’s own words, “never active by default.” It lets an application prove itself with a client secret—a long-lived password the software holds. The same page warns that a leaked secret lets an attacker “silently authenticate as your application until the secret is rotated,” and points to signed JWT assertions—short-lived tokens the app signs with a private key it never transmits—as the safer form. The third escape is the device authorization grant, built for machines that lack a browser: the code appears on one screen, a human types it into another, and the machine polls until approval. The browser is gone. The human is not—just standing somewhere else.
Underneath it all, the standard itself is unfinished. OAuth 2.1 is still an active Internet-Draft—version -16, published 3 September 2026, expiring 7 March 2027—not a finished RFC. Its main grant, the authorization code flow, is redirect-based: it must be started through the owner’s user agent, typically a web browser. The IETF’s new AIMS draft, a working-group document updated 15 September 2026, sets out agent identity by composing existing standards rather than writing an agent-native one, and calls static API keys an antipattern because they “do not convey identity.”
The seam with earlier coverage: the 2% Problem is about whether enterprises switch governance on, the OAuth/SAML bypass wave is about flaws in the layer, Okta’s pre-GA standards market and the forming agent identity stack are about what gets bought. This one is about the flow itself—what actually happens at the moment of authentication.
The login page asks who you are. For an agent, the honest answers are different questions: on whose behalf, and for how long. The page was never shaped to hear either.