Microsoft moved its hosted agent service to general availability this week, giving developers a managed runtime for AI workflows that need hours, days, or weeks to complete. The release, part of Azure AI Foundry, removes the overhead of provisioning and scaling compute for long-running agent orchestration and comes with production guarantees including SLAs and 24/7 support.
The shift from self-managed to fully hosted agent infrastructure
The jump from public preview to GA means hosted agents are no longer experimental. Microsoft now backs the service with standard Azure commitments: 99.9% uptime SLA, enterprise support plans, and integration with the broader security and compliance tooling in Azure Policy and Microsoft Purview. Behind the scenes, Foundry handles compute scaling, state persistence, and automatic restart on failure—crucial for agents that might run through thousands of steps over extended periods.
The service is model-agnostic and framework-agnostic by design. During the preview, Microsoft highlighted integrations with its own Semantic Kernel and AutoGen, but the GA announcement confirms support for any toolchain that can be containerized—LangChain, CrewAI, custom Python orchestrators, and multiple large language models from OpenAI, Mistral, and Meta. Developers deploy an agent as a managed compute endpoint; Foundry then handles message routing, checkpointing, and scaling the underlying resources up or down based on the queue depth.
A newly named capability is the „long-running workflow mode,” which preserves state through a built-in durable execution layer. If an agent calls an external API that takes six hours to respond, the runtime suspends and resumes without developer intervention. This replaces the pattern of hacking together Logic Apps, durable functions, and container instances—a patchwork that often broke under production loads.
What the change means for different teams
For developers, the immediate benefit is removing infrastructure code. Instead of writing and maintaining Kubernetes manifests, VM scale sets, or custom queue-based autoscalers, they define the agent logic and let Foundry manage the rest. This shift lets teams experiment with complex multi-agent patterns—planning, debate, retrieval-augmented generation chains—that were previously too fragile to run unattended.
IT and platform teams get a unified control plane. Agent deployments can be governed through Azure RBAC, network isolation rules, and data residency policies. Cost management dashboards show per-agent and per-workflow spend, so chargebacks become possible. Compliance officers can audit agent behavior through integrated logging to Azure Monitor and Purview, tracking what data an agent accessed and which decisions it made. This level of observability is often missing from homegrown agent deployments.
For business stakeholders, the promise is faster time-to-value for automation projects. A claims-processing agent that must wait for human reviews or external documents can now be built as a single managed service, without a pipeline of half a dozen disparate Azure components. Early adopters in financial services and logistics have reported cutting development cycles from months to weeks for similar pipelines, according to Microsoft’s accompanying case studies.
How we arrived at hosted agents
The road to GA began with the 2023 launch of Azure AI Studio, which morphed into Azure AI Foundry as Microsoft consolidated its portfolio for building and evaluating generative AI. The initial focus was on prompt flow and model benchmarking, but the rapid rise of autonomous agents—fueled by frameworks like AutoGen and LangGraph—exposed a gap. Developers could prototype a multi-agent swarm in a notebook, but pushing it to production required stitching together durable execution, state stores, and observability from scratch.
A private preview in mid-2024 introduced the concept of „hosted agent pool,” offering a turnkey compute target for agent workloads. Over the next six months, Microsoft added support for custom container images, multiple model endpoints, and the long-running workflow mode. At Ignite 2024, the team showed an agent handling a procurement process that ran for three days, interacting with email, ERP systems, and a human approval loop—all on the hosted service. That demo convinced many enterprises the technology was ready for critical workloads.
Parallel to Foundry, Microsoft has been building its own agent ecosystem. Copilot Studio agents for business users sit on top of the same platform, and the company now positions Foundry as the professional-code counterpart for developers and data scientists. The GA of hosted agents solidifies this two-tier approach: low-code for conversational agents, full-code for complex, long-running automations.
Immediate steps to take
If you have agents running on self-hosted Kubernetes or Azure Container Instances, Microsoft recommends migrating to hosted agents to take advantage of SLAs and simplified management. The migration path is straightforward: containerize your agent logic, push the image to Azure Container Registry, and deploy via the Foundry portal or CLI. The same codebase works; only the execution environment changes.
For new projects, start with the documentation’s quickstart template, which sets up a basic long-running agent with state checkpointing in under 15 minutes. From there, you can add framework integrations by selecting them as dependencies in your environment configuration. Microsoft provides pre-built recipes for Semantic Kernel and AutoGen, with LangChain and CrewAI templates expected shortly.
Administrators should establish guardrails now: set up cost alerts on the Azure subscription, define role assignments so only authorized users can create production endpoints, and configure network isolation if agents will access on-premises resources. Review the pricing page carefully—hosted agents bill per core-hour plus a small per-workflow execution fee, and long-running workflows can accumulate cost if not designed with idle timeouts.
Finally, evaluate how hosted agents fit into your broader AI governance framework. The integration with Azure AI Content Safety means you can automatically filter agent inputs and outputs for harmful content, a critical control for customer-facing automations.
What comes next
Microsoft has hinted at several expansions before the end of 2025. Regional availability will broaden beyond the initial US and EU data centers, and the roadmap includes a serverless tier for bursty workloads that only need a few minutes of runtime per execution. Deeper ties with GitHub Copilot are also in the works: developers may soon be able to describe an agent in natural language and have Copilot generate the project scaffold, deployment manifest, and CI/CD pipeline as a single bundle.
Competition is heating up. AWS recently announced its own managed agent service under Bedrock, and Google’s Agent Builder is expanding to support long-running tasks. For Microsoft, the bet is that its hybrid of developer tools, enterprise governance, and AI safety will keep Azure the default choice for organizations already invested in the Microsoft 365 and Power Platform ecosystems. If your team has been building agents in isolation on a virtual machine, this GA is the signal that a better, cheaper way has arrived.