Azure AI Foundry で長時間稼働エージェントの構築が GA にLove this end-to-end example from Jeff Hollan of what is now possible when you build long-running agents with Foundry. All now GA.

Microsoft has moved long-running agent capabilities in Azure AI Foundry to general availability, signaling that developers can now build and operate autonomous, multi-step workflows in production rather than confining them to experiments and previews. The shift matters because most early agent deployments have struggled with tasks that span minutes, hours, or days, where a process must survive interruptions, maintain context, and coordinate multiple tools and models without constant human supervision.

At its core, a long-running agent is designed to handle work that cannot be completed in a single request-and-response cycle. Traditional chat-style assistants operate within a short conversational turn, but many enterprise scenarios, such as processing a large document set, reconciling records across systems, or orchestrating a chain of API calls, require persistence over time. According to Microsoft's description, the feature allows an agent to check in on progress, resume after failures, and continue executing tasks that outlast a typical session. This addresses one of the most cited limitations of first-generation agent frameworks, where a dropped connection or a timeout could force an entire task to restart.

Reaching general availability is significant beyond the technical feature set. GA status typically brings production service-level agreements, formal support, and a commitment to stability that preview offerings do not provide. For enterprise developers, that distinction often determines whether a capability can be used for customer-facing or mission-critical systems. It suggests the underlying orchestration, state management, and reliability mechanisms have matured to a point Microsoft considers dependable for regulated and high-volume environments, although real-world performance will vary by workload.

Azure AI Foundry is Microsoft's unified platform for building, evaluating, and deploying AI applications and agents, and it evolved from what was previously branded Azure AI Studio. It sits alongside a broader Microsoft agent ecosystem that includes Copilot Studio for lower-code business scenarios, the Semantic Kernel and AutoGen developer frameworks for orchestration, and the Azure OpenAI and model catalog services that supply the underlying language models. The long-running agent capability appears intended to give professional developers more control over durable, stateful processes than the packaged Copilot experiences typically expose.

Technically, durable agents of this kind generally rely on mechanisms such as checkpointing, persisted state, and event-driven resumption, concepts familiar from durable workflow engines like Azure Durable Functions. The practical value is that an agent can pause while waiting on an external system, recover from transient errors, and pick up where it left off rather than losing accumulated context. Combined with tool calling, retrieval, and evaluation features already present in Foundry, this is likely to make it easier to compose agents that combine reasoning with concrete actions against enterprise data and services. Microsoft has also emphasized observability and tracing in Foundry, which is important because autonomous multi-step processes are difficult to debug and audit without detailed logs of each decision and action.

The release fits a broader industry push toward agentic AI. Competing platforms have introduced comparable directions, including OpenAI's Assistants and agent tooling, Google's offerings around its Gemini models and Vertex AI, and Amazon's Bedrock Agents. Interoperability efforts such as the Model Context Protocol, which standardizes how agents connect to external tools and data sources, have also gained traction across vendors. Microsoft's move to GA can be read as an attempt to differentiate on operational reliability and enterprise readiness rather than only on model quality, an area where many providers now draw from overlapping foundation models.

For organizations evaluating the capability, several practical considerations remain. Autonomous agents that run for extended periods raise questions about cost management, since sustained model usage and tool invocations can accumulate charges, as well as governance concerns around permissions, data access, and human oversight. Microsoft positions Foundry with identity, security, and content-safety controls through its Azure ecosystem, but the effectiveness of those guardrails in complex agent workflows will depend on how carefully teams design and monitor their deployments. As with any newly generalized feature, adopters will want to validate reliability, latency, and total cost against their specific requirements before committing production workloads.