A futuristic infographic shows a four-step AI workflow with dashboards, cloud computing, human review, and secure servers.
Microsoft used Ignite 2026 on November 17 to launch a "continuous improvement" system for Microsoft Foundry. The company says it is bringing continuous improvement to Foundry that day. Microsoft's pitch is that AI agents shouldn't be shipped once and left alone. Foundry collects traces and expert feedback and turns them into tested changes. Your team decides what reaches production.

The pitch is easy to follow. The details matter more, because some pieces are generally available, some are in preview, and the headline numbers come from Microsoft and its customers. Here is what the announcement contains and what it means for teams running agents.

A futuristic infographic shows a four-step AI workflow with dashboards, cloud computing, human review, and secure servers. What Microsoft announced​

The post comes from Tina Schuchman, the corporate vice president for Microsoft Foundry. It describes four repeating stages: Establish, Identify, Optimize, and Validate. Microsoft places the loop inside its broader Microsoft IQ push. Separate Azure Ignite coverage says the foundation under Microsoft IQ brings Foundry and Fabric together on Azure to build, run, and continuously improve agents.

Constellation Research's Ignite summary frames it in a similar way. It says Foundry and Agent Optimizer will continually improve systems and agents, with quality, speed, cost and outcomes as the real benchmark.

Microsoft's own wording sets the limit on "learning". The system does not rewrite live agents on its own. The post says the optimizer recommends a candidate, and nothing changes in production until your team applies it.

The four-stage loop​

Establish: define what "good" means​

Evaluations are automated tests that score agent responses against criteria you choose. The new pieces here are mostly in preview:

  • Rubric evaluators now cover safety as well as quality (preview).
  • Trace annotations let experts mark what worked, what didn't, and what the agent should have done (preview).
  • Evaluation datasets are built from those annotated real conversations (preview).
  • Per-case rubrics can be generated from those datasets.
  • A golden dataset locks in the best test cases as the benchmark each new version must beat (preview).

The practical gain is that your test set comes from real usage. It does not have to come from hypothetical prompts someone wrote in a meeting.

Identify: find recurring problems​

Insights is generally available. It groups recurring problems found in traces, shows evidence, and points to a likely cause. Microsoft says the same visibility now extends to agents built outside Foundry with LangGraph, LangChain, and CrewAI (GA).

Microsoft Learn's Insights documentation adds some operational detail:

  • Each Insight can include representative traces, affected agent versions, a likely cause, and a recommended next action.
  • The Learn page marks some items as preview. Its documented first analysis uses a 7-day lookback window.
  • Insights supports only GPT models as the judge. Learn recommends GPT-5 or newer, and larger variants give better quality.
  • Learn warns that a high linked-trace count doesn't by itself prove business impact. Teams should review the evidence and the affected workflow.
  • Learn also says the AI-generated findings can be incomplete or wrong and can combine unrelated behavior.
  • Insight generation can incur model charges.

Optimize: candidate changes, tested​

The Agent optimizer improves instructions, skills, and model choice. It tests each candidate against your evaluations. Microsoft says it builds on prompt optimization and the model router, both already available.

New preview capabilities:

  • Sending an Insight to the optimizer, so a production problem becomes a set of tested candidate fixes.
  • Skill optimization, which learns from past work and expert feedback. Skills are versioned, so you can roll back or reuse them in other agents.
  • Cost optimization, which looks for cheaper ways to run an agent, such as a smaller model, without hurting quality.
  • Programmed Training, which lets engineers write Python training recipes that run on serverless GPUs.

Microsoft also renamed managed fine-tuning to Managed Training.

Microsoft Learn explains the mechanics. The optimizer scores a baseline, generates candidates, scores them against the same dataset, and ranks them by a composite score from 0.0 to 1.0. Setup differs by agent type:

  • Prompt agents can be optimized through a portal wizard from the agent's Optimize tab, with no code changes. Learn says this currently excludes agents that use the Copilot harness.
  • Hosted agents need optimizer-ready code and a baseline configuration. Foundry Toolkit can scaffold this, or you can use the Azure Developer CLI.
  • Hosted-agent optimization is available in all regions where hosted agents are available except Norway East. It requires the Responses protocol.

One warning from Learn deserves attention. During a run the optimizer invokes your agent against every task in the dataset. If the agent calls external APIs, databases, or services, those calls really execute. Use test endpoints or mocks to avoid charges, state changes, or rate limiting. Learn also notes that optimized instructions are often longer, which can raise token usage. Check that any score gain justifies the extra cost.

Validate and repeat​

Foundry runs statistical tests when comparing agent versions, so you can tell a real difference from noise. Teams can wire the check into GitHub Actions or Azure DevOps so a release stops if it misses the quality bar. Other pieces:

  • Continuous evaluation samples live traffic with your chosen evaluators. It was already available.
  • Scheduled evaluations rerun the golden dataset.
  • Scheduled red-team scans keep testing for new risks.
  • A per-agent dashboard shows quality, safety, latency, and token usage, with alerts when scores slip.

Microsoft Learn's evaluation overview supports this workflow in broad terms. You can evaluate datasets, generated responses, or production traces, and you can run evaluations during development, in CI/CD, or on a recurring production schedule.

Runtime, identity and platform updates​

The announcement bundles several other items:

  • Entra Agent ID and Foundry guardrails limit what an agent can access. Sensitive actions can keep a human in the loop. Agent controls in Foundry Safety check runtime behavior against your guardrail policies, first for Foundry-built agents and later for agents built elsewhere.
  • Hosted agents handle more concurrent sessions (GA) and start faster (preview).
  • The Microsoft Copilot harness is in preview, with reusable skills and sub-agents.
  • Voice agents are GA.
  • Microsoft 365 accounts for agents are in preview.
  • Hosted-agent deployment to Azure Kubernetes Service is "coming soon" in preview.
  • Cost in Microsoft Foundry is in preview. It gives one view of spend across agents, models, and tools, with a forecast and month-end report.
  • Microsoft-Decision-1.1 is generally available and can act as a judge in Foundry Evaluations.
  • Partner models from fal and Baseten are also available on Foundry.

The numbers, and how to read them​

Every figure below is Microsoft's or a customer's claim. The post does not publish datasets, methodology, or cost calculations.

ClaimSource of claimCaveat
Foundry has more than 11,000 modelsMicrosoftCatalog size says nothing about quality for your workload
Model router saves up to 65% versus sending everything to the frontierMicrosoft's own testingAn upper bound, not a guarantee
Docusign cut AI processing costs by 90% with fine-tuned Foundry modelsMicrosoft, citing DocusignPrior result, not tied to the new loop
Heidi's post-trained open-weight model matched a frontier model on its clinical quality metrics at an estimated 87% lower cost per noteMicrosoft, citing HeidiEstimate, using the customer's own metrics
Alaska Airlines' score rose from 81.2% to 91.1% after tuning MAI-Code-2-FlashMicrosoft, citing Alaska's evaluationQualified employees keep review and final decisions
NTT DATA is validating before scaling to 1,000+ internal projectsMicrosoftStill in validation

On the router claim, Microsoft Learn is more measured. It says savings depend on the workload mix, and workloads with more simple interactions see larger savings. It describes simple interactions as typically 50–60% of traffic. It also lists Quality, Balanced, and Cost routing modes and advises testing changes against representative multi-turn traces. For your own estimate, run your traffic through the router and compare against your baseline.

Timeline context​

This isn't a brand-new idea appearing from nowhere. Microsoft's Foundry developer blog introduced the agent optimizer for Foundry Agent Service in private preview around Build 2026, with public preview planned 30 days later. A third-party guide from daily.dev says the optimizer was announced generally available in October 2026. Microsoft's Ignite post, by contrast, labels several of its new optimizer paths as preview. Treat the GA label as applying to specific features. Check the Learn page for the exact capability you plan to use.

What admins and developers should do​

  1. Build the test set first. Annotated traces and a golden dataset are the foundation. Without representative data, optimization just chases a weak score.
  2. Agree on rubrics with the business. Let domain experts write what "good" looks like, including safety criteria.
  3. Protect trace data. Traces record what agents were asked and how they answered. Check access roles, retention, and privacy rules. The Insights documentation notes that protected tables need specific roles and that data-residency requirements apply.
  4. Budget for the meta-work. Insights, evaluation judges, and optimizer runs all consume model calls.
  5. Isolate side effects. Point evaluation runs at test endpoints.
  6. Gate releases in CI/CD. Use the GitHub Actions or Azure DevOps integration so a candidate must clear your bar.
  7. Keep humans on consequential decisions. Microsoft itself frames the Alaska use case as assisting trained staff.
  8. Check preview terms. Learn says preview features come without a service-level agreement and aren't recommended for production workloads. Regions, supported models, limits, and prices can change.

Analysis​

The strongest idea is the discipline, not the marketing line that agents "get better the more you use them." Production traces feed evaluations, evaluations judge candidate changes, and humans approve them. That is a sensible engineering loop, and it resembles ordinary software release practice applied to prompts, skills, and model choices.

The risks are the usual ones for AI-judged systems. Insights uses AI to infer causes, so its output needs human checking. An optimizer that tunes to a flawed evaluation set will optimize for the wrong thing. The announcement also doesn't spell out pricing, regional availability, data-retention details, or preview limits for the new features. Until those are documented, treat the savings and accuracy figures as marketing, and measure on your own workload.

Microsoft also ties this to its platform. Foundry for running agents, GitHub for building them, Entra for identity, and Azure underneath make for a tidy stack. Teams already invested in Azure will find it convenient. Teams that want portability should note that Microsoft says Insights visibility extends to LangGraph, LangChain, and CrewAI agents, but the optimizer's documented targets are prompt agents and hosted agents.

Bottom line​

Foundry's continuous improvement is a managed pipeline for observing agents, testing changes, and shipping the winners under human control. Several parts are generally available: Insights, voice agents, and more concurrent hosted-agent sessions. Many of the newest pieces are in preview. Pilot it on one agent with a good golden dataset before betting a fleet on it.

 

References

  1. Continuous improvement comes to Microsoft Foundry - Microsoft Azure Microsoft Azure 2026-11-17T18:15:00+00:00
  2. Microsoft Ignite 2026: What's new in Azure, the platform purpose-built for AI azure.microsoft.com
  3. Agent optimizer in Foundry Agent Service overview - Microsoft Foundry | Microsoft Learn learn.microsoft.com