Working Surfacelink · 12 Nov 2026
03
A small boat steered by a clockwork figure waits its turn in a flight of canal locks behind ordinary cargo barges, each lock gate opened by a lock keeper seen in silhouette, the whole scene drawn with wide empty margins on every side.
Snowflake Builders Blog

Agents Are Code Too: What I Learned Shipping a Cortex Agent Through Azure DevOps

Sanket Prabhu · 11 November 2026

How-to · agents · critique & review · Agent ownership

Read the original
Takeaway

Engineer Sanket Prabhu released a finance AI agent on Snowflake through the same versioned, approved pipeline as other production changes, and tested it after every release.

Summary

Sanket Prabhu, an engineer with about ten years of building data platforms, built a finance AI agent with Cortex Agents, a feature of the Snowflake cloud data platform. Building it took a week, but moving it safely through test environments into production took longer. He described the release process on Snowflake's builders blog on Medium.

Key points
  • The agent's instructions, tool descriptions, finance data definitions, search service and access rights all live in one code repository and ship in one release.
  • Releases use the existing Azure DevOps pipeline, Microsoft's software delivery service, with approvals required before the user-testing and production environments.
  • A fixed set of real finance questions runs after every release, checking both the answers and which tool the agent chose to use.
  • Most bad answers in testing came from the finance data definitions, so people who knew those definitions reviewed them in the same pull request.
Implication

An engineering team can release an AI agent like any other production code, with version control, approval gates and a fixed question set run after each release. This is an inference, and the piece is one engineer's account on a vendor's own blog.

Plays
Engineering

After every release, run a fixed set of real questions against the AI agent and check which tool it chose as well as its answers.

Engineering

Keep an AI agent's instructions, tool descriptions, data definitions and access rights in one repository and release them together through the usual approval gates.

Derived by Working Surface from the article; more in the Playbook. Source line: An agent you can’t rebuild from Git is a prototype, no matter how good its answers are.

Source issue

12 November 2026: when AI agents write the code, engineers' main work becomes writing and protecting the checks that decide whether that code is correct.