Microsoft Foundry Adds Continuous Improvement for AI
Summary
Microsoft has introduced continuous improvement capabilities in Microsoft Foundry, enabling organizations to build, run, govern, and iteratively optimize AI agents using real-world traces, evaluations, and expert feedback. The update matters because it helps Azure customers improve agent quality, control cost and performance, and maintain governance as AI workloads scale.
Microsoft Foundry gets continuous improvement for AI agents
Introduction
At Microsoft Ignite, Microsoft announced new continuous improvement capabilities for Microsoft Foundry. For Azure customers building AI agents, this is a notable step because it moves beyond one-time deployment and adds a structured way to improve agents over time using production data, user feedback, and governance controls.
This matters for IT leaders and platform teams that need AI systems to become more accurate, cost-efficient, and trustworthy as usage grows.
What’s new in Microsoft Foundry
Microsoft Foundry now supports an end-to-end improvement loop for AI agents:
- Continuous learning from usage: Foundry can use agent traces and user input to help teams identify issues and improve agent behavior over time.
- Evaluation-driven improvement: Organizations can define what “good” looks like with automated evaluations, rubric evaluators, and golden datasets.
- Insights in GA: The Insights feature is now generally available to surface recurring problems, evidence, and likely root causes.
- Agent optimizer enhancements: Foundry can recommend improvements to prompts, skills, and model selection, then test candidate changes before release.
- New preview features: Microsoft highlighted annotations on traces, evaluation datasets, sending Insights issues to agent optimizer, skill optimization, and cost optimization.
- Large model ecosystem: Foundry supports access to more than 11,000 models across Microsoft, OpenAI, Anthropic, Meta, Mistral, and open-source options.
- Model router savings: Foundry’s model router can send simple requests to smaller models and complex requests to larger ones, with Microsoft claiming up to 65% cost savings versus always using frontier models.
Governance and runtime controls
Microsoft is also emphasizing governance as part of the platform:
- Entra Agent ID gives agents an identity and supports least-privilege access.
- Foundry guardrails and Foundry Safety help enforce runtime policies and keep sensitive actions under human oversight.
- Teams remain in control because every proposed change is tested, reviewed, and explicitly approved before production deployment.
Impact on IT administrators and AI platform teams
For Azure administrators and AI operations teams, the update strengthens Foundry as a managed platform for enterprise agent lifecycle management. It combines development with GitHub, runtime operations on Azure, and governance through Entra and safety controls.
The practical benefit is clearer operational oversight: teams can monitor agent behavior, prioritize recurring issues, tune quality versus cost, and standardize how AI improvements are validated before rollout.
Next steps
If your organization is building or piloting AI agents on Azure, consider these actions:
- Review which Foundry features are GA versus preview before production use.
- Define evaluation criteria and golden datasets for your business use cases.
- Test model router and optimizer features to balance quality, latency, and cost.
- Validate governance settings with Entra Agent ID, guardrails, and human-in-the-loop controls.
Microsoft Foundry is positioning itself as more than an agent runtime. With continuous improvement, it becomes a platform for operating AI agents that can evolve safely and efficiently over time.
Need help with Azure?
Our experts can help you implement and optimize your Microsoft solutions.
Talk to an ExpertStay updated on Microsoft technologies