Microsoft Agent Framework and Durable AI Agents for IT Leaders

In this blog post Microsoft Agent Framework and Durable AI Agents for IT Leaders we will explain why an AI pilot can look impressive in a demonstration but fall apart when it meets a real business process.

The problem is rarely the AI model itself. The problem is that production work takes time, crosses several systems, waits for people and occasionally fails halfway through. If the agent forgets where it was, repeats an action or loses a customer request during a restart, the productivity gain quickly becomes another IT support problem.

A durable AI agent is designed to avoid that outcome. In plain English, it can save its progress, pause when necessary and continue from the right point after a system interruption. That makes it much more suitable for business processes that run for hours, days or even weeks.

What Microsoft Agent Framework actually does

Microsoft Agent Framework is an open-source set of software tools for building AI agents and workflows. An agent is an AI-powered application that can interpret a request, use approved tools and take steps towards an outcome rather than simply producing a written answer.

The framework supports .NET, Python and Go development, although some capabilities differ between languages. It can work with Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude and other model providers, giving organisations more flexibility over the AI models they use.

Workflows add structure around the agent. Instead of telling an AI system to “complete supplier onboarding” and hoping it makes sensible decisions, a workflow can define the required stages, approval points, system connections and recovery rules.

We covered the broader delivery model in what Microsoft Agent Framework means for real-world AI delivery. Durability takes that discussion further by addressing what happens when a workflow does not finish in one clean run.

What makes an AI agent durable

Microsoft combines Agent Framework with its Durable Task technology. Durable Task stores the state of a process outside the individual application instance, coordinates work across computing resources and records which steps have been completed.

Think of it as a save point for a business process. If the application restarts after completing six of ten steps, it can recover its stored state and continue rather than beginning again.

The Durable Extension for Microsoft Agent Framework can provide:

• Persistent sessions so conversation history and process context can survive restarts.
• Workflow checkpoints that record progress at controlled points.
• Automatic recovery when a worker, service or application instance fails.
• Long-running workflows that can pause for an approval, timer or external event.
• Distributed scaling so workloads can be processed across multiple computing resources.
• Monitoring information that helps IT teams understand where a process is waiting or failing.

These agents can run through Azure Functions, which executes code when required without keeping a dedicated server running, or through self-managed infrastructure such as containers and existing applications.

Why durability matters to a mid-market business

Fewer failed processes and less manual cleanup

A traditional AI automation may restart an entire task after a temporary outage. That can create duplicate tickets, repeated emails, inconsistent records or frustrated employees who must work out what happened.

A durable workflow can recover from a known checkpoint. The business outcome is fewer support incidents and less time spent repairing half-completed automations.

Processes can wait for people without losing context

Many important decisions should not be fully automated. A finance manager may need to approve a payment, an HR leader may need to review a recommendation, or a security team may need to approve access.

Agent Framework supports human involvement in workflows. The process can pause, send an approval request and continue when a response arrives, without repeatedly calling an AI model while it waits.

AI can move beyond simple chat

Durability makes more valuable scenarios practical. Examples include customer onboarding, insurance claims, invoice exceptions, IT incident investigation, compliance evidence collection and multi-stage document reviews.

These use cases often involve Microsoft 365, service desks, CRMs, ERPs and custom databases. Our guide to connecting Microsoft Foundry agents to business systems explains why those connections need to be treated as controlled business interfaces, not shortcuts around existing security.

A practical Australian business scenario

Consider a 200-person professional services firm processing new supplier requests. Staff currently copy information from emails, check supporting documents, create a record in the finance platform and chase a manager for approval.

A durable agent workflow could extract the submitted information, check that required documents are present, flag inconsistencies and prepare the supplier record. It could then pause until the finance manager approves or rejects the request.

If approval takes three days or the hosting service restarts overnight, the workflow retains its position. It does not need to reread every document, consume more AI capacity or create a second supplier record.

The business outcome is not “more AI”. It is faster supplier onboarding, fewer data-entry errors, a clearer approval trail and less administrative effort for finance staff.

Durable does not automatically mean safe

This is the most important point for CIOs and CTOs. Durability makes execution more reliable, but it does not make an agent accurate, secure or appropriately authorised.

An agent must still be given a controlled identity and only the permissions needed for its job. High-impact actions should require approval, and external actions must be designed so a retry cannot accidentally create duplicate payments, users or customer communications.

Australian organisations should also assess what personal information the agent collects, stores and sends to models or connected systems. The Australian Privacy Principles remain relevant when AI applications handle personal information.

The Essential Eight, the Australian Government’s baseline cybersecurity framework, does not provide a complete AI governance model. However, its focus on access control, application security, patching, backups and administrative privileges should still influence the environment around the agent.

For a deeper discussion of identities, network controls, monitoring and data boundaries, see designing secure AI agent infrastructure on Azure.

Five questions to ask before investing

1. Does the process genuinely need durability? A short knowledge search may not. A workflow that waits for approvals or coordinates several systems probably does.
2. What happens when a step fails? Define which steps can be retried, which require investigation and how duplicate actions will be prevented.
3. Where must a person remain in control? Payments, account changes, sensitive communications and access decisions are sensible starting points for mandatory approval.
4. How will activity be monitored? Your IT team needs visibility into failures, AI usage, response times, completed actions and unusual behaviour.
5. What business result will justify the cost? Measure processing time, error rates, support effort and turnaround time before and after the pilot.

Start with one process, not an enterprise-wide agent

The safest approach is to select one repetitive process with a clear owner, measurable cost and manageable risk. Begin with read-only access, add human approval and only then consider allowing the agent to make controlled changes.

Microsoft Agent Framework gives development teams useful building blocks, but successful delivery still depends on architecture, security and process design. The technical implementation should support a defined business outcome rather than becoming an experiment looking for a problem.

CloudProInc brings more than 20 years of enterprise IT experience to this work as a Microsoft Partner and Wiz Security Integrator. Our Melbourne-based team works hands-on across Azure, Microsoft 365, OpenAI, Claude, Microsoft Defender and Wiz rather than handing clients to a large, disconnected delivery team.

If you are considering a durable agent but are unsure whether the process, security model or likely return justifies the investment, we are happy to review the idea with you — no strings attached.