Snowflake adds background agent runs to its managed MCP server
Snowflake announced on November 4 that its managed MCP server now supports background runs for CORTEX_AGENT_RUN and CODING_AGENT tools. A new background boolean argument controls the behavior. When it is set to true, the tool returns immediately with a handle that the client can poll.
The release describes asynchronous execution rather than a guarantee that the underlying agent finishes faster. A client can receive a handle before the work completes. The announcement does not provide a latency benchmark, a completion-time commitment or a new service-level agreement.
Snowflake automatically adds three supplementary tools to an MCP server that contains at least one CORTEX_AGENT_RUN or CODING_AGENT tool. ASYNC_AGENT_RUN_STATUS lets a client poll a run using its handle. ASYNC_AGENT_RUN_CANCEL lets it request cancellation using that handle.
The third supplementary tool, AGENT_RUN_RESULT, reads the reply of an agent run using a run handle. These additions separate starting work, checking its state and reading the result. The release does not say that every existing MCP tool becomes asynchronous.
The same announcement adds a content_cel_filter parameter to CORTEX_AGENT_RUN and CODING_AGENT. Snowflake says a client can request an outline first and then read only the blocks it needs. The note does not quantify token savings or establish that filtering eliminates all unwanted content.
For teams integrating agents, the stated change is a handle-based way to start, inspect, cancel and retrieve supported runs. The release does not announce a new price, an exemption from account permissions or a durability guarantee. Those limits should remain separate from the newly documented tool behavior.