Foundry Agent Service is generally available (GA). Some sub-features are in public preview and might have different constraints.
Prerequisites
- An Azure subscription.
- A Microsoft Foundry project.
- A deployed model compatible with Agent Service. Model and region availability can vary.
Supported regions
The following table shows regional support for the Responses API, agents, and voice-based agents (preview). It also shows support for private Class A IP address ranges. For voice-based agents, the region must support Agent Service and the required Voice Live capabilities for the selected model and channel.
Some Azure OpenAI models aren’t available in every region. For details, see Region availability for Foundry Models sold by Azure.
Not all tools are available in every region. For example, file search isn’t available in Italy North and Brazil South. For the full tool-by-region matrix, see Tool support by region and model.
Regional support for private networking
When you use a private network configuration, such as a network-secured standard agent, the following regional requirements apply:- Foundry resource and virtual network region. You must deploy the Foundry resource in the same region as its virtual network. You can deploy other Azure resources, such as Azure Cosmos DB, Azure AI Search, and Azure Storage, in different regions. Consider the cost implications of cross-region deployments.
- Grounding with Bing Search. Only the following regions are supported: West Europe, Canada East, Switzerland North, Spain Central, UAE North, Korea Central, Poland Central, Southeast Asia, West US, West US 2, West US 3, East US, East US 2, Central US, South India, Japan East, UK South, France Central, Norway East, Australia East, Canada Central, Sweden Central, South Africa North, Italy North, Brazil South.
Supported models
Agent Service supports Azure OpenAI models and several Foundry models sold by Azure. Model availability can change over time and varies by region. To see the full list of models you can deploy for your project and region, use the Foundry portal model experience.Sovereign clouds
Foundry Agent Service is also available in Azure Government (US Gov Virginia and US Gov Arizona) with a subset of agent types and tools. For the full list of supported features, see Foundry Agent Service feature availability in Azure Government.Tool support by region and model
Region and model determine which tools are available to your agent. In the following tables, Yes means fully supported, No means not supported, and Limited means partial support that varies by tool configuration. Check individual tool documentation for details. The following table shows which tools are available in each supported region. This table only accounts for service availability, so make sure the model you want to use is also available in the same region. The following table shows which tools each model supports. For the image generation tool, you need both thegpt-image-1 model and a large language model (LLM) as the orchestrator in the same Microsoft Foundry project.
Troubleshooting
A model or version isn’t available in your region
- Confirm you selected the right tab for your deployment type (global standard vs. provisioned).
- Try a different region that supports the model and Responses API.
- If you’re using gpt-5 models, registration is required. Access is granted according to Microsoft’s eligibility criteria.
A tool isn’t available in your region
- Not all tools are supported in every region. For example, file search isn’t available in Italy North and Brazil South.
- Check the tool support by region and model table to confirm availability before you deploy.
- If a tool isn’t available, choose a supported region or use a different tool.
Provisioned throughput deployment fails
- Confirm you have enough PTUs available in the region.
- Review Provisioned throughput and Spillover traffic management.
Agent receives rate-limit (429) errors
- If the error code is
session_quota_exceeded, stop sessions that don’t need active compute, delete sessions whose state is no longer needed, or request a limit increase. - If the error code is
regional_session_quota_exceeded, retry with exponential backoff or use another region. If the error persists create an Azure support request. - Implement exponential backoff with jitter in your application retry logic.
- For sustained high-throughput workloads, consider provisioned throughput deployments.
- Review Azure OpenAI quotas and limits for your deployment’s tokens-per-minute and requests-per-minute caps.
How Agent Service enforces limits
Foundry Agent Service enforces limits in two places:- Agent Service limits. Limits for agent and thread artifacts, such as file uploads, vector store attachments, message counts, and tool registration.
- Model limits. Quotas and rate limits for the model deployments your agents call.
Where Agent Service stores your data
Where your agent data lives depends on which setup option you choose. The setup option also determines which resources the quotas and limits in this article apply to.- Basic setup stores agent state in secure, Microsoft-managed storage that’s logically separated. This is the default when you don’t configure your own resources.
- Standard setup stores agent state in customer-managed, single-tenant Azure resources in your own subscription, which gives you full control over data residency and access.
Capability settings tell Agent Service where to store and process file uploads, vector stores, and conversation history. To store agent data in your own resources, see Standard agent setup.
Foundry Agent Service endpoints are regional, and data is stored in the same region as the endpoint. For more information, see the Azure data residency documentation.
Default service limits
Agent Service applies a per-subscription limit to concurrent hosted agent sessions in each region. This limit includes sessions across all Foundry accounts and projects in the subscription and region. A session counts toward the quota while its compute is being provisioned or is running. An idle or stopped session retains its persisted state but doesn’t count toward the quota. Resuming it requires available quota.
These values are the default limits. To request a higher concurrent session limit, create an Azure support request. In the request, specify the subscription, region, and expected number of concurrent sessions. Limit increases depend on the available regional capacity.
The following table lists default limits enforced by the Agent Service. These limits apply to all Foundry projects regardless of subscription type or region.
The Agent Service limits in this table are fixed and apply uniformly across all subscription types. Rate limiting for model calls is applied at the model deployment level; see Azure OpenAI quotas and limits for model-specific rate limits.
Voice-based agents can also incur voice-session, audio, channel, or Voice Live limits.
Limit error reference
When you exceed a limit, the Agent Service returns an error. Handle these errors gracefully in your application.
For example:
- File exceeds the maximum size. Uploading the file fails. Split the content into smaller files or reduce file size before you upload.
- Vector store token limit. Attaching a file to a vector store fails if the file exceeds the token limit. Reduce the file content or split it into multiple files.
- Thread message cap. Adding messages can fail after a thread reaches the message limit. Create a new thread for a new conversation, or archive and rotate threads as part of your application design.
- Message content size. Creating a message can fail if the
textcontent is too large. Send smaller messages, or move large content into files and use file search. - Tool registration cap. Creating or updating an agent can fail if you register too many tools. Register only the tools you need, and prefer fewer, reusable tools.
- Rate limit exceeded. API calls to the model deployment are throttled. Implement exponential backoff with jitter.
- Concurrent session quota exceeded. Creating or resuming a hosted agent session fails when the subscription reaches its concurrent session limit in the region. Delete unused sessions or request a limit increase.
- Regional session capacity exceeded. Creating or resuming a hosted agent session fails when the region has insufficient capacity. Retry with exponential backoff or use another region.
- Concurrent session quota exceeded. Creating or resuming a hosted agent session fails when the subscription reaches its concurrent session limit in the region. Stop sessions that don’t need active compute, delete sessions whose persisted state is no longer needed, or request a limit increase.
- Regional session capacity exceeded. Creating or resuming a hosted agent session fails when the region has insufficient capacity. Retry with exponential backoff or use another region.
- Delegated subnet has no available IP addresses. Creating or resuming a hosted agent session fails when the delegated subnet is out of addresses. Stop sessions that don’t need active compute, or configure a larger subnet before retrying.
- Agent version cap. Creating a version fails once an agent reaches 1,000 versions. The service returns HTTP 400 with the message
Maximum number of agent versions (1000) exceeded. Please delete older versions before creating new ones.This error is terminal, not a transient one, so don’t retry the same call. Delete versions you no longer need to free capacity immediately, then create the new version.
Best practices to stay within limits
Use the following practices to reduce limit-related failures:- Keep files small and focused. Prefer multiple smaller documents over a single large document.
- Avoid very large messages. Put long content in uploaded files and query it by using file search.
- Plan for long conversations. Treat threads as conversation history and rotate to new threads when conversations become very long.
- Register only required tools. Remove unused tools from agent definitions.
- Manage agent versions. Each agent allows up to 1,000 versions. Delete or rotate versions you no longer need as part of your deployment pipeline to stay well under the cap, and don’t retry create calls that fail with the version-cap error until you free capacity.
- Monitor usage trends. Track agent activity by using Foundry Agent Service metrics to identify growth before you hit limits.
Model quotas and rate limits
Agents follow the quotas and rate limits for the model deployments they use. For current model quotas and limits, see: To view or request more model quota, see Manage and increase quotas for resources with Microsoft Foundry (Foundry projects).Request a limit increase
The limits in this article are default values for Foundry Agent Service. If your workload requires higher limits:- Model quotas. You can request increases for model deployment quotas. See Manage and increase quotas for resources with Microsoft Foundry.
- Hosted agent session quotas. To request a higher concurrent session limit, create an Azure support request. Specify the subscription, region, and expected number of concurrent sessions.
- Agent Service limits. The file, message, and tool limits listed in this article are fixed service limits and can’t be increased. Design your application to work within these constraints by using the best practices described earlier.