Microsoft’s new Copilot Autopilot experience is built on OpenClaw, according to the two companies’ announcements. Microsoft described Autopilot as a proactive agent entering an initial customer preview; OpenClaw says Microsoft’s engineering contributions include policy-conformance checks, native Windows work, sandbox integration, and reliability improvements that were contributed upstream.
This is notable both as a large-scale product built on an open agent framework and as a concrete example of upstream work flowing back into that framework. It does not mean that every feature in Microsoft’s product is available in OpenClaw, or that a default OpenClaw installation inherits Microsoft’s product-level controls.
From Scout to Copilot Autopilot
Microsoft’s September 25 Copilot announcement introduced Home, Code, and Autopilot as parts of its Copilot experience. It described Autopilot as a persistent, proactive personal agent with its own identity, memory, and workspace, and said an initial customer preview was expected at the end of September.
Microsoft had introduced the underlying product as Scout in June and said it was powered by OpenClaw. The September announcement uses the name Autopilot. This is Microsoft’s Copilot feature—not Windows Autopilot, the separate device-provisioning product with a similar name.
The rollout status matters: Microsoft’s post described a preview, not general availability. Eligibility, regions, supported accounts, and final capabilities can change as the preview expands.
What OpenClaw says moved upstream
OpenClaw’s September 25 account says Microsoft’s contributors worked on several layers of the integration:
- Policy conformance: checks for providers, network and MCP access, secrets and authentication, and message routing.
- Windows product experience: native WinUI chat and approval flows.
- Execution isolation: a backend for Microsoft Execution Containers (MXC).
- Runtime reliability: scheduler and duplicate-reply handling, loop safeguards, memory retention, and other operational fixes.
Those are OpenClaw’s descriptions of contributions and should be read as project reporting, not as an exhaustive Microsoft security specification. The public announcements do not establish that every contribution is enabled in every OpenClaw environment or that the integration has identical behavior outside the Copilot preview.
Why upstream contributions matter to agent builders
Agents that can use tools, reach network services, or communicate across channels need more than a capable model. They need an explicit policy for which providers, MCP servers, credentials, destinations, and actions are allowed. They also need isolation boundaries, human approval for consequential actions, careful secret handling, and operational protections against loops and duplicate messages.
When enterprise integration work improves reusable framework components, other operators may benefit from reviewable changes in the upstream project. The practical check is to inspect the actual release notes, configuration defaults, and deployment documentation before assuming a safeguard applies to your version or installation.
What this does not prove
Microsoft’s product announcement does not certify every self-hosted OpenClaw setup, nor does it mean OpenClaw can access a user’s Microsoft data without the product’s own account, permission, and rollout controls. The available posts also do not provide a complete threat model, formal security audit, or feature-by-feature comparison between Autopilot and the upstream project.
For an independent deployment, restrict tools and credentials to the minimum required, isolate execution, log administrative actions, review MCP and network access, and keep a human approval step for writes that affect customers, money, or production systems. Test upgrades and recovery paths before expanding access.
FindMilan’s OpenClaw setup service helps teams plan installation and safety controls. The Hermes agent setup project is a separate agent deployment example, not Microsoft’s Autopilot implementation.
Official sources
Frequently asked questions
Is Microsoft Copilot Autopilot built on OpenClaw?
Microsoft's September 25, 2026 Copilot announcement names Autopilot as a new Copilot experience, and Microsoft's June 2 Scout announcement—the product's earlier name—said Scout was powered by OpenClaw. OpenClaw's September 25 post describes the upstream contributions and integration work.
Is Microsoft Copilot Autopilot generally available?
Microsoft described an initial customer preview for the end of September 2026. The announcement does not establish general availability; check Microsoft's current rollout and eligibility information.
What did Microsoft contribute to OpenClaw?
OpenClaw's account describes policy-conformance checks, native Windows interface and approval work, a Microsoft Execution Containers sandbox backend, and scheduler and runtime reliability improvements. The details should be attributed to OpenClaw's report rather than treated as a complete Microsoft product specification.
Does the Microsoft integration make every OpenClaw installation secure by default?
No. Product-specific policy controls and sandbox integrations do not secure every self-hosted deployment automatically. Operators still need to configure permissions, secrets, network access, channels, logging, and isolation for their own threat model.