Setting Up AI Solutions in Foundry

• Foundry defines three assignable scopes: the resource, the project, and an individual agent
• Connections are created at account level for shared dependencies and at project level for the resources that back a project's agents
• Capability hosts enable Agent Service on the account and project, and they cannot be updated after they are set
• Basic agent setup uses Microsoft-managed multitenant storage; standard setup pins agent state into your own single-tenant Cosmos DB, Storage, and AI Search
• Pay-per-token standard deployments need no compute quota and bill on tokens; managed compute reserves accelerators and bills on uptime
• Data at rest stays in the designated geography for every deployment type; the deployment type decides only where inferencing data is processed
• Provisioned deployments buy reserved capacity in PTUs for lower and more consistent latency; standard deployments are best-effort
• Batch deployment types trade real-time responsiveness for roughly half the cost and a separate enqueued-token quota
• Instant access lets you call supported models by name with no deployment created at all
• A standard deployment's version upgrade policy decides whether it follows the new default, waits for expiry, or stops serving at retirement
• Which deployment options exist is a property of the chosen model, and partner or community models add an Azure Marketplace purchase step
• Prompt agents are declarative definitions the service runs; hosted agents are your own container image on Microsoft-managed infrastructure
• Agents are identified by name plus version, and creating a version is what provisions a hosted agent's runtime
• Structured inputs override fixed tool configuration per request without creating a new agent version
• Publishing an agent requires at least Foundry Project Manager at the Foundry resource scope
• Standard agent setup provisions several Cosmos DB containers with per-container throughput requirements, and undersizing fails capability host provisioning
• Standard setup requires account-level control-plane roles plus granular data-plane roles on each dependency for the project managed identity
• Outbound isolation injects the agent client into a delegated subnet, and that outbound configuration cannot be changed after deployment
• Private endpoints to the bring-your-own Search, Storage, and Cosmos DB dependencies are not created for you
• Govern AI workloads with one inherited baseline policy assignment plus narrower scoped assignments and exemptions, never duplicated policy sets or a loosened global control
• Retry with backoff only covers transient faults; surviving a dependency outage needs a cached last-known-good copy plus a circuit breaker
• azd pipeline config wires the repository to Azure and GitHub Actions signs in with OpenID Connect rather than a stored secret
• Foundry evaluation runs as an automated quality gate inside the pipeline and fails the job when scores fall below a threshold
• Bicep and ARM templates provision the whole Foundry footprint, including existing resource IDs for bring-your-own dependencies
• Multi-agent releases validate version compatibility and roll out progressively with rollback triggered by quality-regression metrics
• IaC sets desired state at deploy time; a deployIfNotExists policy plus a remediation task is what re-fixes resources that drifted afterwards

Unlock with Premium — includes all practice exams and the complete study guide.