git:20260519.7b28ce1 to git:20260520.ff62120

15 added, 16 removed. Audit A to A.

---

name: preset-cortex-agents

- description: Use Snowflake Cortex Agent REST and SQL APIs through direct API/operator workflows. Use only for direct API workflows when a user asks to list, describe, create, update, delete, or run Cortex Agent objects; call ad hoc agent:run; handle streaming Cortex Agent responses; or use the SNOWFLAKE.CORTEX.DATA_AGENT_RUN SQL wrapper. Do not use for MCP-only work.

+ description: Use Snowflake Cortex Agent REST and SQL APIs for listing, describing, creating, updating, deleting, running agents, streaming responses, and SQL wrappers. Use only for direct API workflows; Do not use for MCP-only work.

---

# preset-cortex-agents

- Use this skill for Snowflake Cortex Agent object and run workflows.

+ Use for Snowflake Cortex Agent object management and run workflows.

- ## Workflow

+ ## Always

- 1. Use `preset-snowflake-cortex` first to establish account URL, auth, role,

- warehouse, database, schema, and safety context.

- 2. Load [references/agent-runs.md](references/agent-runs.md) before object-based or ad hoc `agent:run` calls.

- 3. Load [references/agent-management.md](references/agent-management.md) before listing, describing, creating, updating, or deleting Cortex Agent objects.

- 4. Load [references/sql-agent-ddl.md](references/sql-agent-ddl.md) when the user wants to create, alter, show, describe, or drop Cortex Agent objects from SQL.

- 5. Load [references/sql-wrapper.md](references/sql-wrapper.md) when the user wants to run an agent from SQL instead of REST.

- 6. Load [../preset-snowflake-cortex/references/cortex-safety.md](../preset-snowflake-cortex/references/cortex-safety.md), summarize the target, payload, tools, budgets, expected data exposure, and rollback path when applicable, then get explicit confirmation before any run or mutation.

+ - Use `preset-snowflake-cortex` first for account, auth, role, warehouse, database/schema, privileges, and safety context.

+ - Prefer read-only list/describe before mutations.

+ - Require confirmation of query, role, tools, budget, and output handling before any run.

+ - Require rollback planning before create, update, replace, or delete.

+ - Treat streamed events as sensitive; store redacted summaries unless raw output handling is approved.

+ - Do not rely on pre-2025 Cortex Agent schemas.

- ## Guardrails

+ ## Retrieve

- - Prefer read-only list/describe calls before mutations.

- - Do not run agents without explicit confirmation of the query, role, tools, budget, and output handling.

- - Do not create, update, replace, or delete agent objects without a rollback plan.

- - Do not assume old `agent:run` schemas are current; Snowflake updated the request and response schemas in 2025, so do not rely on pre-2025 documentation or examples.

- - Treat streamed events as potentially sensitive. Store only redacted summaries unless the user approves raw output handling.

+ - Object-based or ad hoc `agent:run`: [references/agent-runs.md](references/agent-runs.md)

+ - List/describe/create/update/delete agents: [references/agent-management.md](references/agent-management.md)

+ - SQL create/alter/show/describe/drop: [references/sql-agent-ddl.md](references/sql-agent-ddl.md)

+ - `SNOWFLAKE.CORTEX.DATA_AGENT_RUN` SQL wrapper: [references/sql-wrapper.md](references/sql-wrapper.md)

+ - Safety and approval checklist: [../preset-snowflake-cortex/references/cortex-safety.md](../preset-snowflake-cortex/references/cortex-safety.md)