Microsoft has quietly assembled the plumbing for a persistent, stateful agentic AI platform inside Azure, and the implications for enterprise developers and independent software vendors are more significant than any single product announcement suggests.
For the past two years, Copilot looked like a chat window bolted onto Microsoft 365. That framing is now obsolete. Through a combination of the Durable Task Framework, the newly announced Microsoft Agent Framework, and a May 2026 Copilot Studio release that introduced computer-using agents and a redesigned visual workflow canvas, Microsoft has systematically rebuilt its AI stack from the substrate up. The goal is not another assistant. It is a durable execution layer that can run multi-step, multi-agent pipelines across enterprise environments for hours, days, or weeks without losing state.
The Durable Task Extension for Microsoft Agent Framework is where the technical story gets interesting. Built on top of Azure Durable Functions, it provides checkpointing and replay semantics for AI workflows, meaning that if an agent is midway through a 200-step orchestration and the infrastructure hiccups, the job resumes from exactly where it left off rather than starting over. According to Temporal's Series D announcement, the company raised $300 million led by Andreessen Horowitz in February 2026, reporting 380 percent year-over-year revenue growth and 9.1 trillion lifetime action executions, at a $5 billion valuation. The market is clearly willing to pay for durable execution at scale, and Microsoft is now offering a version of it as a native Azure capability rather than a separate vendor relationship.
The Microsoft Agent Framework itself merges two previously separate projects: AutoGen, the multi-agent research system out of Microsoft Research, and Semantic Kernel, the enterprise-oriented SDK that has been Microsoft's primary developer surface for AI integration. Combining them into a single commercial-grade runtime is a significant consolidation. For ISVs and enterprise developers already building on Azure, the message is straightforward: you no longer need to stitch together a separate orchestration layer. The connectors ship pre-built for Azure AI Foundry, Microsoft Graph, Fabric, SharePoint, and dozens of SaaS systems via Azure Logic Apps.
The honest comparison to make is not between Microsoft and Salesforce but between Microsoft's stack and the open-source orchestration ecosystem. LangGraph reached v1.0 in late 2025 and is now the default runtime for all LangChain agents, running in production at Klarna, Uber, and LinkedIn. Its model is graph-based reasoning with checkpoint-backed state between nodes. Temporal operates differently, providing event-history-backed durability at the workflow level rather than within individual reasoning steps. The practical architecture that many production teams have landed on is Temporal for macro-level workflow orchestration and LangGraph for micro-level agent reasoning inside each step.
Microsoft's Durable Task approach attempts to collapse that two-layer pattern into a single Azure-native surface. Whether it succeeds depends on how deeply developers are already committed to the LangGraph or Temporal ecosystems, and whether Microsoft's execution guarantees prove as robust in practice as Temporal's replay semantics. The framework is still in public preview, which means the production maturity question is genuinely open.
The Salesforce Agentforce comparison is more about market positioning than architecture. Agentforce 3 with its Atlas reasoning engine is winning on multi-step customer-facing service workflows, particularly in CRM-heavy environments. Copilot Studio is winning on document and email-heavy internal workflows, which makes sense given the 20 million paid Microsoft 365 Copilot enterprise seats as of April 2026, with more than 60 percent of Fortune 500 companies running at least 10,000 seats. Most large enterprises will likely operate both platforms for different functions, meaning the real competition is for developer mindshare and the ISV ecosystem built on top of each.
The lock-in question every enterprise developer should be asking
Infrastructure bets compound. When Microsoft ships pre-built connectors for its own data services, teams that build agentic workflows on Azure naturally wire those workflows into Fabric, Graph, and SharePoint. Each integration makes a future migration more expensive. This is not a criticism so much as an honest description of how platform gravity works. Microsoft's scale advantage, 300 million-plus Microsoft 365 commercial seats globally, means the distribution channel for these agents is already embedded in the daily workflow of most enterprise knowledge workers.
The more granular risk for ISVs is the question of differentiation. If Microsoft ships durable orchestration, built-in connectors, governance controls, and multi-agent coordination as platform primitives, the surface area for differentiated middleware products shrinks. That is the same dynamic Temporal and LangGraph are navigating from the open-source side, leaning into model-agnosticism and portability as their counter to platform gravity.
Microsoft Build 2026, scheduled for June 2 and 3 in San Francisco, is expected to bring more announcements around the Windows Agent Framework APIs and a Windows Agent Store, extending the agentic surface area from cloud to desktop. The Copilot Tasks feature, already in rollout, lets scheduled agents execute autonomously against a user's calendar and email without manual triggering. The infrastructure layer is converging, and enterprises that have been treating Copilot as a chat add-on are about to face a considerably more consequential decision.
Also read: Federal lawmakers move to ban emotion-reading AI as the science behind it crumbles • Andrew Kelley banned AI code from Zig and the reasoning is harder to dismiss than most critics expect • MiniMax chose Hong Kong over Wall Street and its shareholders are glad it did
