A business unit wants weekly updates to a customer-support Microsoft Foundry agent, but change windows are limited and you must avoid downtime while…

The scenario requires frequent updates with zero downtime, strict latency SLOs, and measurable ROI for each release. A canary release strategy is the industry standard for these requirements. It routes a small percentage of live traffic to the new agent version, allowing for performance monitoring (latency, ROI) against the baseline. If metrics are met, traffic is gradually increased. If they fail, an automatic rollback can be triggered, minimizing impact and meeting the no-downtime constraint. This approach provides the necessary control and data for rollback decisions.

A. Big-bang production swap with extended monitoring and manual rollback: This is high-risk, can cause significant downtime if issues arise, and makes isolating the impact of the new release for ROI measurement difficult.

B. Duplicate agent per business unit: This creates management overhead and architectural complexity without solving the core problem of safe, iterative releases for a single agent.

C. Freeze models; update prompts only: This fails to meet the business requirement for weekly updates to the agent, which may include model or logic changes, not just prompts.

1. Microsoft, "Safe rollout for online endpoints," Azure Machine Learning Documentation. This document details practices like blue-green and canary deployments (traffic mirroring and gradual rollout) to safely deploy new model versions without downtime, which directly aligns with the question's constraints. (Section: "Safe rollout strategies")

2. Le-Cun, Y., & Haffner, P. (1999). "Object Recognition with Gradient-Based Learning". In Shape, Contour and Grouping in Computer Vision. Springer. While a foundational text, it establishes the principle of iterative model improvement, which modern MLOps practices like canary releases are designed to manage safely in production. (Concept of iterative updates).

3. Microsoft, "Design a CI/CD pipeline," Azure Architecture Center. This guide outlines best practices for continuous integration and deployment, emphasizing progressive exposure techniques like canary releases to reduce risk and validate changes against production traffic. (Section: "Release pipeline")