Developer proposes formal class/instance definition of AI agents against 2026 MCP spec
A developer building an MCP host has proposed an operational definition of what an 'agent' is, arguing that neither the A2A nor MCP protocols formally define the term. The definition splits the concept into two parts: an agent definition (the class) covering the host loop, system context, tool policies, LLM, and termination limits, and an agent instance holding the goal, working context, credentials, and consumed budget. The author notes that popular framings like Anthropic's 'LLM using tools in a loop' are directionally correct but lack the precision needed to build a host around. The article also highlights that the 2026-07-28 MCP spec removes protocol sessions entirely, requiring all state to travel in explicit handles visible to the model. The proposed structure is stress-tested against the latest MCP specification, which was set to finalize on July 28, 2026.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in