What's New in Snowflake Cortex Agents for Enterprise AI

What's New in Snowflake Cortex Agents for Enterprise AI

Skip to content

Blog / AI & ML / Your Agents Are in Production. Now What?

JUL 21, 2026 / 10 min read AI & ML Your Agents Are in Production. Now What?

Arun Agarwal +3

The first generation of enterprise AI focused on building agents. The next generation is about operating them at scale.

Cortex Agents have always provided a managed runtime for building AI applications inside Snowflake. Over the past year, we've watched customers move from deploying their first agent to managing enterprise deployments. That shift has surfaced a new set of operational challenges around orchestration, execution and governance.

Today, we're expanding Cortex Agents with new capabilities that simplify the operational complexity of building, deploying, and operating enterprise AI agents at scale.

At a glance

Build Coding Agent (in public preview soon): Build a managed coding agent backed by the same runtime that powers CoCo and deploy it in any application.

Skills Package (in public preview soon): Package a curated set of skills into a single object and reference them in your agent with a single URI

Agent Toolset (in public preview soon): Reference another agent's tools in your current agent specification instead of rewriting and maintaining duplicate tool definitions.

Tool Search (in public preview soon): Discover tools on demand instead of loading every definition upfront.

Run Async Agent API (generally available soon): Run long jobs in the background without holding a connection open.

Code Execution Tool (in public preview soon): Process data and generate polished outputs inside a sandboxed Python environment.

Interrupt and Resume (generally available soon): Stop a running agent, correct course and pick up where it left off.

Partial Access (in public preview soon): Serve users with different permission levels from a single managed agent.

Manage Versioning UI (in public preview): Compare, roll back and promote agent configurations without editing YAML.

Cortex Agents — a refresher

A Cortex Agent can reason over your enterprise data, invoke tools, execute code and take actions across connected systems — all within Snowflake's governance boundary. Every action is subject to the same access controls and audit trail that cover the rest of your data.

You define a Cortex Agent declaratively that includes its orchestration model, tools and instructions. Applications invoke it through a REST API, while Snowflake handles model routing, tool invocation and result assembly.

The production challenge

Most deployments start with a single agent, a handful of tools and short tasks. That setup works fine for a proof of concept and maybe a small deployment.

Enterprise production at scale looks different: Workflows span multiple specialized agents, tool libraries grow to dozens of integrations, requests run for minutes rather than milliseconds and governance teams need visibility into exactly what each agent is authorized to do. These aren't problems that better prompting or a faster model will solve. They're operational problems, and they're what our new capabilities below are designed to address.

Build your agent

Bring Snowflake CoCo's coding capabilities to any application

Snowflake CoCo is the first truly Data-native AI coding agent, built to speed time to value across modern data stack. Deeply aware of your enterprise's data, compute, governance, and operational semantics, CoCo turns complex data engineering, analytics, machine learning, and agent-building tasks into simple and informed conversations with high accuracy and trust.

With the Cortex Agents REST API, you can now build a managed agent that brings those same coding capabilities (in public preview soon) into your own applications and interfaces. Configure the agent with the CoCo toolset (including sandbox), scope its instructions and data grounding to your environment, and expose it over a standard REST endpoint. Any query that requires code generation, SQL execution, data transformation or pipeline automation is handled inside the sandboxed runtime — the same foundation that powers CoCo, now composable as part of your own agentic workflows.

For example, our engineering team built "Debug with CoCo," an internal tool that root-causes any Cortex Agent request, in under 20 min. The app itself is a few hundred lines long: it takes a request_id via UI, and makes a single Cortex Agent API call with a dedicated prompt. The managed API provisions a CoCo sandbox, mounts the logs as the agent's grounding, runs bash and SQL through the CoCo toolset, and streams an evidenced root-cause analysis back to the browser. No coding-agent runtime, and no data leaving Snowflake — just REST calls:

Figure 1: "Debug with CoCo," an engineer tool built on the Managed Agent API with CoCo runtime.

Give your agents a shared capability layer

As agent deployments grow, skill management becomes a bottleneck. Every new agent needs its own list of capabilities, every update needs to be applied across each one individually, and as skill libraries expand, agent specifications become harder to maintain and easier to get out of sync.

Skills Package (in public preview soon) lets you define a curated set of skills once and reference the entire collection in any agent specification with a single URI. When the package is updated, every agent that references it picks up the change automatically. Teams can build a shared capability layer — a standard set of data workflows, domain-specific tools, or approved integrations — and compose new agents on top of it without duplicating or maintaining separate copies.

Reuse tools across agents without rewriting them

Enterprise agent deployments often start with a well-defined agent — one with a carefully curated set of tools, tested integrations and tuned configurations. When a second agent needs some of those same tools, the current approach is to duplicate the tool definitions: copy the spec, maintain two copies and keep them in sync whenever something changes.

Agent Toolset (in public preview soon) lets you reference another agent's tools directly in your current agent's specification. A single line points to the source agent, and all its tools are available immediately — no redefinition required. When the source agent's tools are updated, every agent referencing that toolset inherits the change automatically. Teams can define a canonical set of tools once — a shared...