Concept card

Snowflake Cortex Agents

As of 2026-07-24T14:00:00Z

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is Snowflake Cortex Agents?

Cortex Agents chains Cortex Search, Cortex Analyst, and SQL/Python tasks in one agentic loop using a tool-call protocol compatible with Anthropic and OpenAI conventions, hosted on Mistral Large 2 or Arctic Instruct by default.

What it is

Snowflake Cortex Agents is Snowflake's agentic orchestration layer: a runtime that sits above individual Cortex functions and lets a language model decide, at each turn, which tool to call next — Cortex Search for unstructured retrieval, Cortex Analyst for structured natural-language-to-SQL, a Python or SQL task, or a write-back to a table — rather than executing a single fixed query. For an SAP analytics consultant increasingly asked to advise on multi-platform estates (SAP alongside Snowflake, Databricks, or Fabric), Cortex Agents is the piece that explains what "agentic AI" concretely means on the Snowflake side of a client's landscape, and it is the natural comparison point when scoping SAP's own agent stack (Joule Studio, BDC agent orchestration) against a competing platform.

Why it matters

The decision heuristic is simple: if a human analyst would need more than one application to answer the question, you need Cortex Agents, not Cortex Analyst alone.
Every tool call executes inside Snowflake's compute boundary, so results and any write-back stay governed rather than leaving the platform.
Claude and GPT-4o are available as optional external model routes, not the default — worth checking before assuming a specific model is in play.

Key points

Cortex Agents = Snowflake's multi-step agentic AI runtime: tool composition + retrieval + write-path, built above Cortex Analyst.
Architecture: reasoning model + tool set (Python/YAML) + Cortex Search index → directed tool-calling loop → structured output via Snowflake Tasks.
Cortex Analyst is callable as a sub-tool — NL query becomes one step in a multi-step agent workflow.
Write path requires a procedural wrapper (Task/Stored Procedure) — the agent does not write to tables directly.
Pricing: credit-based, materially higher than Cortex Analyst — a 5-step, 3-tool, 4 k-token workflow ≈ 10–15 credits at list.
BDC pattern: Joule agents handle SAP business-semantic grounding; Cortex Agents handle Snowflake-native workloads; the BDC Knowledge Graph links the two surfaces.
Failure modes: hallucinated tool calls, write-path race conditions, context window overflow — each needs explicit mitigation by design.
Best fit: Snowflake-native workflows requiring data + reasoning + action; avoid when the primary data is SAP-semantic (use Joule instead).

Terms used on this page

Cortex Analyst: Snowflake's natural-language-to-SQL interface — a single-step tool that translates a business question into a SQL query and returns the result. Cortex Agents can call Cortex Analyst as one tool in a multi-step workflow.
Cortex Search: Snowflake's managed vector search service — indexes documents or unstructured text and returns semantically relevant chunks, used in Cortex Agents for retrieval-augmented generation (RAG).
Directed tool-calling loop: The agent architecture where the reasoning model iteratively selects a tool, receives its output, decides whether to call another tool or synthesise a final answer, and repeats until the task is complete.
Snowflake Task: A Snowflake serverless scheduler that runs SQL or Stored Procedure code on a defined trigger — the recommended write-path mechanism for Cortex Agents outputs that need to persist to a table.
Arctic: Snowflake's own open-weight foundation model, optimised for enterprise SQL and data tasks — the default reasoning model in Cortex Agents when no external model is specified.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →