Build with GitHub, contextualize with Microsoft IQ, run in Foundry, govern with Agent 365.
These four building blocks form the core of Microsoft’s “agent platform”—at least the one it presented at Build 2026.
Here we revisit this framework, focusing on a few features that have recently moved into commercial phase or were launched in experimental phase.
A desktop app for GitHub Copilot
Over the past weeks, the GitHub Copilot app has entered technical pre-release. It is available for Windows, Mac and Linux (x64, Arm). There is a waiting list for users of the free version of GitHub Copilot. Not for subscribers of Pro, Pro+, Max, Business and Enterprise.
GA mid-June for Work IQ, the context engineering layer
Under the Work IQ banner, Microsoft claims a unified context layer, from data to tools.
Beyond access to data from its software suites (Office, as well as Dynamics 365 and PowerApps via Dataverse) and external systems via Copilot connectors, Microsoft promises to go beyond anchoring by “understanding how organizations work” (skills profiles, major projects, communication habits…).
The endpoints (A2A, MCP and REST) will move to commercial phase on June 16. Billing will be consumption-based, independent of Microsoft 365 Copilot licenses—though these licenses are necessary for certain features. More precisely, there will be variable fees for query-type usages (retrieval, reasoning) and fixed fees for invoking actions and tools. Microsoft provides three examples, of increasing complexity:
- Identify tasks or actions assigned by a manager and turn them into a check-list : $0.20 to $0.40
- Analyze the latest customer emails to identify the main themes and the impact on the roadmap, then recommend 3 actions: $0.30 to $0.75
- Produce Level-1 and Level-2 summaries of the most recent roadmap review, drawing on meetings and priority documents for the CMO: $0.50 to $1.50
Web IQ, a “premium” version of Bing-based anchoring
Complementing Work IQ, there is Web IQ. This suite of APIs, available in preview (waiting list), is a form of a “premium” version of Bing anchoring. It also relies on the search index (augmented by licensed data), but has its own architecture for vectorization, ranking, extraction and routing.
Microsoft has integrated in this pipeline a few of its technologies. For example, its embedding Harrier model and its DiskANN nearest-neighbor search algorithm. It claims a latency of 164 ms at the 95e percentile.
Frontier Tuning, a managed reinforcement learning environment
Another feature in preview, as part of the Forward Deployed Engineers initiative launched with EY: Frontier Tuning. It provides access to a managed reinforcement learning environment. Microsoft intends to integrate it into Copilot Studio and Foundry.
Half a dozen in-house LLMs, including a flagship at 1T
There are also developments in Microsoft’s LLM catalog. The flagship is called MAI-Thinking-1. It is an MoE (35 billion active parameters out of about 1,000 billion), with a 256k context window. Microsoft compares its performance to Claude Opus 4.6 on SWE-Bench Pro (coding). It also pits it against Claude Sonnet 4.6 on user preference (1276 tasks unspecified).
In the coding category there is a specialized model: MAI-Code-1-Flash. Microsoft compares it notably to Claude Haiku 4.5. It has been deployed in GitHub Copilot within VS Code.
Foundry, in connection with Fireworks AI
Other models added to the catalog: those from Fireworks AI, within Foundry. On the menu: DeepSeek, Google, Meta, Mistral AI, Moonshot AI, Qwen and Zhipu AI.
The integration has moved to GA. It also enables inference on custom models. Possible bases: Kimi (K2, 2.5 and 2.6), GLM (4.7 and 4.8), OpenAI (gpt-OSS-120b), Qwen (3.5-9B, 35B-A3B, 112B-A10B and 397B).
MDASH, a response to Claude Mythos
Another GA: the integration between Defender and GitHub Code Security (part of the GitHub Advanced Security suite). Requiring Defender CSPM to be enabled, it enriches the vulnerabilities detected in code with runtime context.
MDASH, meanwhile, is still in preview. (accessible to members of the Security Advisors program). It represents Microsoft’s answer to Claude Mythos: an “agentic security system” that orchestrates “more than 100 specialized agents” to detect vulnerabilities. It has helped uncover 16 — including 4 RCEs — on Windows’ network/authentication stack (Microsoft fixed them in its latest Patch Tuesday). About fifty partners have joined the initiative. Including, in France, Atos, Capgemini and Genetec.
MXC, the prospect of a composable sandbox to run agents
Another preview on the security side: MXC (Microsoft eXecution Container). The idea is to provide a sandbox that is “composable.” It is capable, on Windows, Mac and Linux, of leveraging several isolation back-ends, behind a unified configuration scheme and a TypeScript SDK.
MXC is meant to strengthen the “agentic control plane” Agent 365. Microsoft has integrated it into the GitHub CLI. NVIDIA is leveraging it to port OpenShell to Windows. OpenClaw also uses it, to run a node and gateway.
