GCP Agent Shadow Deployer

Implements the "Dark Canary" pattern for Cloud Run, allowing agents to be evaluated in production without serving user traffic.

GoogleCloudPlatform Updated 1.1k repo stars

File contents

gcp-agent-shadow-deployer

This skill helps you implement safe, automated rollouts for your AI agents. Based on the "Evaluating Agent Performance" codelab, it uses Cloud Run revision tagging to deploy a Shadow Revision for evaluation before promoting it to live traffic. This enables System-Level Evaluation (latency, network, auth) rather than just unit testing.

Usage

Ask Antigravity to:

  • "Deploy a shadow revision of my agent for evaluation"
  • "Tag my new Cloud Run revision with the Git SHA"
  • "Promote my shadow revision to 100% traffic"

Shadow Pattern ("Dark Canary")

  1. Tag with Commit Hash: Every deployment is tagged with the current Git commit hash (e.g., c-a1b2c3d).
  2. --no-traffic: The revision is deployed with 0% traffic.
  3. Shadow URL: Access the revision via a dedicated tag URL: https://[TAG]---[SERVICE]-[HASH]-[REGION].a.run.app.
  4. Promotion: If evaluation metrics (Final Response Match, Tool Use Quality) pass, promote the revision to 100% traffic.

Shell Template

Refer to scripts/deploy_shadow.sh for the standard implementation.

googlecloudplatform/devrel-demos commit 723e18f765

Frequently asked questions

npx get-skill add googlecloudplatform/gcp-agent-shadow-deployer