Microsoft Work IQ Adds Durable A2A and MCP Tasks With Progress and Cancellation

Microsoft is planning to address that exact operational headache for developers building on its workplace intelligence layer. According to Microsoft's AI at Work roadmap (Roadmap ID 570851), an upcoming update titled "Work IQ: A2A and MCP long running tasks" is currently in development for Work IQ and Microsoft Copilot (Microsoft 365). Targeted at the Developer platform for Worldwide (Standard Multi-Tenant) cloud instances, the feature is slated for Preview in October 2026 and General Availability in November 2026.

While a single-sentence roadmap entry can easily fly under the radar, this specific update targets one of the most important plumbing problems in modern software engineering: giving both Agent-to-Agent (A2A) and Model Context Protocol (MCP) integrations durable, long-running task execution complete with progress tracking, delayed result retrieval, and explicit task cancellation.

The Core Update: Roadmap ID 570851 at a Glance​

Before diving into the architectural implications, it helps to separate what Microsoft has placed on the schedule from what is already running in production. Because Microsoft's roadmap entries represent estimated targets rather than guaranteed ship dates, enterprise architects should treat these milestones as planning horizons:

The crucial takeaway here is the four-part lifecycle promised in the roadmap description: durability, progress tracking, later result retrieval, and cancellation, applied across both A2A and MCP pathways.

What Is Work IQ, and Why Do A2A and MCP Matter?​

To understand why durable task management is such a pivotal upgrade, we first need to look at how Work IQ operates today. According to Microsoft Learn's Work IQ documentation, Work IQ serves as the workplace intelligence layer that lets AI agents securely access, ground, and reason over organizational data across Microsoft 365—including emails, calendar events, OneDrive and SharePoint files, Microsoft Teams messages, Planner plans, and people context—as well as business and workflow data from Dynamics 365 and Power Platform.

Rather than forcing enterprise developers to build custom retrieval-augmented generation (RAG) pipelines, manage vector databases, or manually replicate compliance rules, Work IQ exposes a unified intelligence layer through three primary protocol surfaces:

1. Agent-to-Agent (A2A): Designed for autonomous multi-agent collaboration and delegation. If you have an operations agent or a Microsoft Copilot Studio agent that wants to hand off a complex workplace question to Work IQ and receive a processed outcome, A2A is the delegation protocol of choice.
2. Model Context Protocol (MCP): Available as both a remote MCP server and a local MCP server (via the Microsoft Work IQ CLI). MCP exposes Microsoft 365 context and actions as a compact surface of 10 generic, verb-based tools (such as fetching, creating, or updating resources) for LLMs, IDEs, and coding assistants.
3. REST API: A standard conversational request-response interface tailored for web applications and backend orchestrators.

The Architectural Divide: Delegation vs. Tool Invocation​

Why is it noteworthy that Roadmap ID 570851 brings long-running task controls to both A2A and MCP at the same time? Because under the hood, A2A and MCP represent two fundamentally different philosophies of agent orchestration.

As documented in Microsoft's Copilot Developer Camp labs and Microsoft Learn's Work IQ API overview, A2A treats Work IQ as a peer agent. Your calling agent discovers Work IQ via its agent card ([url]https://workiq.svc.cloud.microsoft/a2a/.well-known/agent-card.json[/url]) and communicates over JSON-RPC at [url]https://workiq.svc.cloud.microsoft/a2a/[/url], using the A2A-Version: 1.0 header for v1.0 methods like SendMessage (or defaulting to v0.3 if omitted). In this model, the calling agent delegates a high-level goal, and Work IQ returns a structured task object containing a task ID, a contextId for multi-turn continuity, a status state, and generated artifacts.

MCP, by contrast, is traditionally a tool-invocation protocol where the calling LLM or orchestrator remains in the driver's seat, calling stateless tools to pull context or execute actions. Industry analyses from TrueFoundry and developer commentary from Coding with Ramin and Matthew S. Wong highlight how the broader AI ecosystem has wrestled with state and long-running work across these two standards throughout 2026—with A2A centering on server-owned task objects and MCP evolving task extensions to prevent long jobs from being chained to brittle, open HTTP connections.

Section Summary: Work IQ acts as a permission-trimmed bridge to Microsoft 365, Dynamics 365, and Power Platform data. While A2A is built for agent-to-agent task delegation and MCP is built for LLM tool invocation, complex enterprise workloads frequently push both protocols beyond simple, instantaneous request-response cycles.

From Real-Time Streaming to True Task Durability​

If you have already experimented with Work IQ's A2A interface, you might ask a fair rhetorical question: Doesn't Work IQ already support long-running operations?

In a partial sense, yes—and comparing what exists today with what Roadmap ID 570851 describes reveals the true value of the upcoming update.

According to Microsoft's Copilot Developer Camp documentation for the Work IQ A2A protocol, Work IQ's agent card currently advertises "streaming": true (while "pushNotifications" and "stateTransitionHistory" are listed as false in the lab's example card). When a developer uses the SendStreamingMessage and SubscribeToTask operations over A2A, the client opens an active streaming channel to receive real-time chunks and generated artifacts—such as asking Work IQ to analyze 10 days of upcoming meetings, suggest preparation topics, and generate a formatted Word (.docx) document saved to OneDrive for Business. Similarly, on the MCP side, the Work IQ CLI package has previously introduced basic progress notifications for supported MCP clients during lengthy queries.

However, based on general distributed systems engineering principles and the contrast between existing documentation and the new roadmap description, real-time streaming is not the same thing as durable task management:

• The Fragility of Active Streams: Streaming requires the calling client or orchestrator to hold a connection open while the work happens. If a mobile client loses connectivity, an IDE restarts, or a serverless orchestrator times out waiting for a massive cross-tenant search and document synthesis job to finish, a purely stream-bound workflow can orphan the work.
• Later Result Retrieval: By explicitly introducing durable long-running tasks with later result retrieval, the roadmap item points toward an asynchronous pattern where a calling agent or MCP host can kick off a heavy job, disconnect or move on to other tasks, and poll or fetch the completed output later.
• Progress Tracking Across Protocols: Instead of flying blind while Work IQ reasons across mail, Teams, and SharePoint, both A2A and MCP consumers are slated to get structured visibility into how a long-running task is progressing.
• The Cancellation Kill-Switch: In enterprise environments, stopping an unneeded job is just as important as starting one. Whether a user realizes they asked the wrong question or an automated orchestrator hits a timeout threshold, native task cancellation prevents wasted compute and unnecessary API consumption.

(Note: While Microsoft Learn's Work IQ overview separately documents a feature called "Workspaces"—which uses SharePoint Embedded working storage to give agents persistent space within the Microsoft 365 tenant boundary for intermediate data and multi-agent handoffs—Microsoft has not stated whether Roadmap ID 570851's durable A2A/MCP task lifecycle uses Workspaces under the hood. Until Microsoft publishes the technical specifications for 570851, developers should treat Workspaces and the new A2A/MCP task lifecycle controls as distinct architectural concepts.)

Section Summary: While Work IQ already supports synchronous calls and real-time streaming over A2A, Roadmap ID 570851 introduces asynchronous durability—allowing A2A and MCP clients to track progress, walk away and retrieve results later, or cancel active jobs outright.

Security, Governance, and the Copilot Credits Factor​

Why should IT administrators and FinOps teams care about a developer-facing protocol enhancement? Because in the Work IQ ecosystem, agent execution is tightly bound to user identity, enterprise compliance, and usage-based billing.

According to Microsoft Learn's Work IQ documentation, every Work IQ API request operates under strict security and licensing rules:

• Delegated User Context Only: Work IQ requires Microsoft Entra ID delegated authentication (supporting On-Behalf-Of flows for multi-agent chaining). Application-only (daemon) authentication is not supported. Every tool call or A2A task runs strictly in the context of the signed-in user, enforcing Microsoft 365 permissions, sensitivity labels, and compliance policies via a centralized, Rego-based policy engine.
• Multitenant Token Matching: For complex parent/child multitenant organizations, Microsoft Learn warns that the access token's issuer (iss) must match the signed-in user's home tenant authority (AzureADMultipleOrgs), or requests will fail with a 400 AuthenticationError.
• Usage-Based Billing via Copilot Credits: Work IQ API access is independent of standalone Microsoft 365 Copilot seat licensing and is billed on a consumption basis using Copilot Credits, with governance and cost management controls housed in the Microsoft 365 admin center. Even for users who hold a Microsoft 365 Copilot license, custom and third-party agent invocations of Work IQ are subject to usage-based billing.

When you view task cancellation and later result retrieval through the lens of usage-based billing, the practical benefits come into sharp focus. From an industry analysis perspective, if a custom agent triggers a sprawling, multi-step research task that goes off the rails, having a native protocol mechanism to cancel the task—and having durable retrieval so a client doesn't have to re-run an expensive prompt just because a network socket dropped—is a major win for cost control and operational hygiene.

Open Questions and How Engineering Teams Should Prepare​

Transparency requires pointing out what Microsoft's roadmap entry doesn't tell us yet. Because Roadmap ID 570851 is still in development and detailed API reference updates have not yet been published for the new lifecycle methods, several technical questions remain open:

• Exact Wire Mechanics: What specific JSON-RPC methods or MCP protocol extensions will govern polling, result retrieval, and cancellation?
• Retention and Timeouts: How long will a completed task's output remain available for "later result retrieval" before expiring, and what are the maximum execution timeouts for a durable task?
• Cancellation Semantics: When a cancellation request is issued mid-flight, how quickly does execution halt, and how are partial actions or Copilot Credit charges handled?

Even with those implementation details pending, Windows and Microsoft 365 developers building agentic workflows in Copilot Studio, Azure, or custom IDEs can take several practical steps today:

1. Audit Timeout Assumptions in Agent Orchestrators: Review your existing A2A and MCP client implementations to identify where synchronous HTTP timeouts or brittle streaming loops are limiting the complexity of prompts you can send to Work IQ.
2. Verify Entra ID Delegated Auth Flows: Because durable tasks allow results to be retrieved later, ensure your token caching and On-Behalf-Of (OBO) refresh flows in Microsoft Entra ID are properly configured to handle asynchronous follow-up calls without failing authentication.
3. Watch for the October 2026 Preview: Keep an eye on the Microsoft Learn Work IQ API documentation and the @microsoft/workiq CLI release notes as the October 2026 preview window unfolds, followed by the targeted November 2026 general availability rollout.

Agentic AI is rapidly outgrowing the era of simple chat bubbles. By bringing durable task execution, progress tracking, delayed retrieval, and cancellation to both A2A and MCP, Microsoft is laying the groundwork for Work IQ agents that can tackle heavy, asynchronous enterprise work without leaving developers—or their users—staring endlessly at a spinning wheel.