None of the 98,291 Registered MCP Tools Expose the Capabilities Needed to Rule Out External-Effect Anomalies
This paper builds an effect-history model separating events in the external world from the runtime's observations of them, then catalogs eight recurring anomalies at the agent-tool boundary under retries, speculative execution, concurrency and partial failures: missing effects, duplicated effects, aborted effects that survive, and committed effects depending on provisional state later withdrawn. Measuring the standard annotation vocabulary across 98,291 tools exposed by registered MCP servers, it finds the fields are widely emitted but give only coarse call-level hints, and none of the required boundary capabilities is expressible. Four of the eight anomalies cannot be excluded by black-box tool invocation at all.
↳ Follow the thread