Unverified50% confidenceOpinionExact time
MCP only addresses how capabilities are described and invoked, not permission narrowing, auditing, or cost control, which remain the harness's own responsibility.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Is the Agent Harness Dead? A Cursor Engineer's Deep Dive
bilibiliMaiYangAI10/21/2026
Related Entities
Related Claims
VerifiedMCP's permission model is relatively coarse-grained: once a tool is registered, the agent can invoke it, lacking a context-aware risk assessment mechanism for individual calls, built-in operation rate limiting, or mandatory human approval processes.78% similarVerifiedMCP only addresses the standardization of connectivity and invocation, and is not responsible for security; responsibilities such as authentication and authorization, approval of dangerous operations, and verification of service trustworthiness must be handled by the application layer.77% similarUnverifiedMCP only specifies "how capabilities are described and invoked," not how the model makes decisions, so it solves interface-layer problems rather than an agent's planning and reasoning capabilities.76% similarUnverifiedReal control comes from the permission grant point—that is, the capability set exposed by the MCP server, plus hard spending caps and kill switches that are independent of the prompt; a prompt-level "please be careful" does not constitute control.76% similarUnverifiedIn MCP architecture, authorization decisions should not remain solely at the authentication layer but need to be pushed down to the actual execution stage of tool calls.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1081672API
curl https://kongchang.com/api/v1/knowledge/claims/1081672MCP
get_claim(id=1081672)