
A stalled Microsoft Foundry project still has its resources, its agents and its model deployments. The team that chose them may have moved on, and Microsoft may have changed the platform underneath them since. Before anyone changes code, you need to know what is there and which of those changes stand between the project and production.
The week-one audit answers five questions, in this order. Where does the project live? Which agent API does its code call? Are its models still served? Can anyone measure its answers? Who can reach it? Each answer narrows the next. The project type decides which portal shows the project and which software development kit (SDK) its code needs. The SDK then tells you which agent API it calls. The result is one list of findings, ranked by what blocks production, and the fixes work through that list. Microsoft Foundry is the current name of the platform that was Azure AI Foundry and, before that, Azure AI Studio (per Microsoft's guide to migrating from the Foundry (classic) portal). An inherited project's documentation may use any of the three names.
Start in the Azure portal, in the project's resource group. Microsoft's current resource model is one Foundry resource with child projects under it. The earlier model combined a hub with an Azure OpenAI resource and an Azure AI Services resource (per What is Microsoft Foundry? on Microsoft Learn). Whatever the name on the project, the Azure resource type stays Microsoft.CognitiveServices/accounts. The current Foundry portal shows only Foundry projects. A hub-based project therefore looks missing until you switch to the Foundry (classic) portal (per the classic portal migration guide).
If the resource group holds a hub, the project is hub-based. Microsoft's guide to moving from hub-based to Foundry projects lists what transfers to a new Foundry project on the same Foundry resource. Model deployments, data files, fine-tuned models and vector stores transfer. Preview agent state (messages, threads and files) does not, and neither do open-source model deployments. The hub-based project also has no access to the new projects. The same guide says Foundry projects do not yet have every feature that hub-based projects have. Record which features the project uses before you plan a move.
Note the Foundry resource's region as well. The Responses API and Foundry Agent Service are not available in every Azure region. A Foundry resource in an unsupported region cannot create or run agents in the current Foundry portal (per the classic portal migration guide).
Microsoft set August 26, 2026 as the sunset of the Assistants API and the retirement of the azure-ai-inference package (per Microsoft's classic portal migration guide). Agents written for the Assistants API work in threads, messages and runs, and are created with create_agent(). Current agents are built on the Responses API. They work in conversations, items and responses, and are created as agent versions with create_version(). Search the code for those calls and for the AzureOpenAI() client. Also check the version of the azure-ai-projects package, because 1.x targets the classic portal and 2.x the current one. Microsoft warns that a 2.x sample run against a 1.x setup causes errors.
Microsoft's guide to migrating to the new Foundry Agent Service sets four steps at minimum. Compare the current tools with what the new agents support. Migrate the agent definitions. Replace threads with conversations and runs with responses. Then verify state, tool calls, outputs and error handling. Microsoft's migration tool rewrites code such as agent definitions and run creation, but not state data. Past threads, runs and messages do not carry over. Put each agent on the findings list with the API it calls and the tools it uses.
A model deployment that answered when the project stalled can stop answering. Microsoft's Foundry Models lifecycle and support policy moves each model through five stages: preview, general availability, legacy, deprecated and retired. Once a model version is retired, every inference request to it returns the error 410 Gone. Under the standard lifecycle, a generally available model's retirement date is set at launch, 18 months out. New customers lose access when it is deprecated at 12 months. List every deployment with its model, version and deployment type. Then check each against Microsoft's model retirement schedule or the Models API.
Read the Models API with care, because its labels differ from the portal's. Its lifecycleStatus value Deprecating means what the portal calls deprecated, which still serves existing customers. Its value Deprecated means retired (per the Foundry Models lifecycle and support policy).
What happens at retirement depends on the deployment type, as the table below sets out from the same policy. A deployment whose versionUpgradeOption is set to NoAutoUpgrade is never upgraded and stops working at retirement. Microsoft's own example is gpt-4o version 2024-05-13, which retires on January 14, 2027, with gpt-5.1 as its replacement (per the Foundry Models lifecycle and support policy).
| Global Standard, Data Zone Standard and Standard | Provisioned | |
|---|---|---|
| At retirement | to the replacement automatically, region by region | not upgraded |
| Quota | handles the quota for those upgrades | make sure quota is available for the target model |
Microsoft emails retirement notices to the subscription owners with active deployments, at least 60 days before a generally available model retires. It also posts them as Azure Service Health advisories filed under Azure OpenAI Service (per the Foundry Models lifecycle and support policy). On an inherited subscription, check who those owners are today.
Microsoft Foundry groups observability into three capabilities (per Microsoft's overview of observability in generative AI). Evaluation scores the quality, safety and reliability of model and agent outputs. Monitoring tracks token use, latency, error rates and quality scores in production through Azure Monitor Application Insights. Tracing records model calls, tool calls and agent decisions. In the current Foundry portal, past evaluation runs sit under Build, then Evaluations. Traces sit under Operate, then Tracing (per the classic portal migration guide).
A project with neither has no record of how well it answered before it stalled, so a fix has nothing to be measured against. Before you change anything, build a small evaluation dataset from the questions the agent was built to answer. Run it with Foundry's built-in evaluators, such as groundedness, relevance, tool call accuracy and task completion. Groundedness checks whether an answer is supported by its source material. Microsoft's pre-production evaluation measures task adherence, groundedness, relevance and safety before an agent goes live (per the observability overview). Week one gives you the same baseline for the version you inherited.
Foundry separates control plane access, for managing resources, from data plane access, for development. Owner or Contributor alone is not enough for some development work, such as chatting with an agent (per Microsoft's guide to rolling out Microsoft Foundry across an organization). The Foundry roles were renamed, with the same role IDs. Azure AI User, Azure AI Owner, Azure AI Account Owner and Azure AI Project Manager became Foundry User, Foundry Owner, Foundry Account Owner and Foundry Project Manager. An inherited role assignment may carry either name (per the classic portal migration guide). List who holds each role and which connections still authenticate with API keys. The rollout guide prefers Microsoft Entra ID and keeps API keys for cases where fine-grained roles are not needed.
Then check how the project is isolated. When one Foundry resource holds several projects, they all share the damage from production incidents, exhausted quota and misconfiguration. Microsoft's advice for production is to treat isolation as the default, with one Foundry resource per production workload boundary (per the rollout guide). A stalled project that shares its Foundry resource with experiments goes on the findings list before it goes near production.
By the end of the week you have one list, ranked by what blocks production. Blockers come first: a retired or retiring model with no replacement deployment or with upgrades set to NoAutoUpgrade, agent code still written for the Assistants API, a Foundry resource in a region without the Responses API. Risks come next: a hub-based project the current portal does not show, no evaluation baseline, production sharing a Foundry resource with experiments, API keys where Entra ID roles belong. Housekeeping comes last. The fixes then work down the list, and the evaluation dataset from week one shows whether each fix helped.
Ovidius takes over Microsoft Foundry projects another team started, as part of AI managed services. We start with a technical assessment of what is built. Then we fix what we find, make the agent production-ready and run it in weekly sprints. Nothing changes in production without passing through staging first, and the Foundry deployments run inside your Azure tenant.
Check where the project lives before anything else. The resource group shows whether it is a Foundry project on a Foundry resource or a hub-based project. That decides which portal shows it and which SDK its code needs. Then check which agent API the code calls, whether each model deployment is still served, whether evaluations and traces exist, and who holds which role. Rank what you find by what blocks production.
The current Foundry portal shows only Foundry projects. A hub-based project appears only in the Foundry (classic) portal, as Microsoft's classic portal migration guide says. To move it, you create a new Foundry project on the same Foundry resource. Model deployments, data files, fine-tuned models and vector stores transfer. Preview agent state, open-source model deployments and hub project access do not.
The Assistants API reaches its sunset on August 26, 2026, according to Microsoft's classic portal migration guide. Agents built on it are rewritten for the Responses API in Foundry Agent Service. Threads become conversations, runs become responses, and agents are created as agent versions. Microsoft's migration tool rewrites code such as agent definitions and run creation. Past threads, runs and messages do not carry over.
The retired version stops answering. Every inference request to a retired model version returns 410 Gone, under Microsoft's Foundry Models lifecycle and support policy. Global Standard, Data Zone Standard and Standard deployments are upgraded to the replacement automatically. If versionUpgradeOption is set to NoAutoUpgrade, the deployment stops working instead. Provisioned deployments are not upgraded, so you create the replacement deployment and move the traffic yourself.
Yes. Run an evaluation before you change anything, so each fix has a baseline to be measured against. Foundry's built-in evaluators score outputs for groundedness, relevance, tool call accuracy and task completion, among others. You run them on a dataset of the questions the agent was built to answer. Tracing records model calls, tool calls and agent decisions, as Microsoft's observability overview describes.
Yes. Ovidius starts with a technical assessment of what is built. We then fix what we find, make the agent production-ready and run it in weekly sprints as part of AI managed services. Nothing changes in production without passing through staging first. The Foundry deployments run inside your Azure tenant, as the Microsoft Foundry and AI managed services pages describe.
Inherited a stalled Foundry project? Book a discovery call. A senior partner runs the call and tells you whether Foundry is the right platform for the job before anything is scoped.