Microsoft launches Copilot Autopilot on OpenClaw
Microsoft launched Copilot Autopilot, which it describes as an always-on mode for delegated, ongoing work in its rebuilt Copilot app. Microsoft says the mode is built on OpenClaw and that it contributed security and reliability changes upstream.

TL;DR
- Copilot is now organized around Home, Code and Autopilot, as yusuf_i_mehdi's announcement describes, with Chat, Cowork and in-app Office work concentrated in Home.
- Autopilot is the renamed Scout, according to steipete's Scout clarification, and Microsoft describes it as a cloud-hosted agent with a tenant-bound identity, memory, computer and workspace.
- Microsoft built Autopilot on OpenClaw, steipete said, while openclaw says Microsoft engineering work is flowing back upstream on security and reliability.
- The rollout has started through Frontier, yusuf_i_mehdi's rollout reply says, but Autopilot itself expands to private preview at month-end and its long-running work is usage billed.
Jared Spataro, Microsoft's chief marketing officer for AI at Work, says in the launch announcement that Autopilot can watch channels, run recurring work and resume a project days later. OpenClaw's contribution recap lists a policy-conformance plugin, message-routing checks, sandbox integrations and guards against repeated tool calls among the work being shared upstream.
Home and Code
Home combines the quick-turn Chat interface with Cowork, where a defined task runs end to end. Code turns a natural-language request into an app, tracker, dashboard, automation or workflow.
- Home: The launch announcement says Chat handles questions, lookups and drafts, while Cowork returns completed artifacts such as an RFP response or customer briefing.
- Office in Copilot: Microsoft says Word, Excel and PowerPoint files are live, editable documents rather than generated exports, with changes synchronized between Copilot and the Office apps.
- Code: The same announcement says Code uses the same underlying technology as GitHub Copilot, runs in a sandboxed environment and can be hosted within the customer's tenant.
Managed Runtime
Code, Cowork and Copilot Studio applications share Microsoft Copilot Managed Runtime, now in public preview. The Managed Runtime announcement describes a Microsoft-operated host inside the Microsoft 365 tenant boundary.
Its control plane includes:
- Microsoft Entra identity and sharing.
- Policies for connectors, enterprise data, approved endpoints and auditing.
- Deployment, versioning and lifecycle controls, plus a central Microsoft 365 admin-center inventory for access, usage and health.
- An SDK and CLI that scaffold projects, configure data connections, generate typed TypeScript services, preview apps and deploy versioned builds.
Microsoft says the host already powers Cowork, Code and Copilot Studio, and the SDK opens the same runtime to third-party tools and professional developers.
Autopilot identity and workspace
Autopilot was previously called Scout, as steipete's Scout clarification confirms. Microsoft presents it as a persistent personal agent rather than a task that expires with the chat session.
Microsoft says in the launch announcement that an Autopilot receives a name, role and goal, then watches channels, follows threads, handles recurring work and resumes paused projects. Its cloud-hosted workspace has its own identity, memory and computer within the tenant.
The product is exposed in Teams, Outlook, chats, channels and documents, where people can @mention it. Microsoft says permissions, audit and governance sit behind that access, while the creator supplies the objective and boundaries.
OpenClaw upstream
Microsoft's OpenClaw work predates this Copilot redesign. steipete said the teams had worked since March to prepare the codebase for large-scale deployments.
OpenClaw's contribution recap attributes several changes to Microsoft engineers:
- A policy-conformance plugin for security posture, model and provider, network and configuration checks, plus authentication and secret handling.
- Message-routing checks intended to ensure a conversation reaches the right agent.
- Windows-native chat, approvals, sandbox flows and execution-container integration.
- Scheduling, background processing, restart and long-conversation recovery work, plus guardrails against repeated tool use.
Microsoft's June account of Scout's origins traces the effort to Omar Shahine's Lobster project, which began as three OpenClaw agents with separately scoped security profiles and tool access. The September deployment does not necessarily track OpenClaw mainline exactly: steipete characterized Microsoft's version as an extended-stable release.
Frontier and usage-based billing
The umbrella rollout began through Frontier, yusuf_i_mehdi's rollout reply says. Microsoft's more specific schedule places Home and Code in Frontier over the coming weeks, Code in broad availability in the following weeks and Autopilot's private-preview expansion at the end of the month; Code is slated for a Microsoft 365 Premium and Pro preview later this year.
Copilot's cost boundary is split by work type. The launch announcement says a user subscription license covers everyday Chat and Office work, with Auto routing each request by accuracy, speed and cost. Cowork, Code, Autopilot, long-running agentic capabilities, Astra and Fable instead run under usage-based billing.
The launch post does not name a model-by-model routing policy for Auto. emollick criticized the use of unspecified models behind a router, arguing that routers can underestimate task difficulty and yield poor outputs.
Test audit
Separate from the Copilot rollout, OpenClaw said it deleted about 400,000 lines of its own tests with little change in code coverage. steipete's test audit tied the cleanup to a project skill meant to identify low-value tests generated by models.
OpenClaw's test-audit skill gates a proposed test on four questions:
- What observable behavior does it test?
- What contract does it independently enforce?
- Why would existing coverage miss the failure?
- Does it require a production seam that no caller needs?
Campaign mode prunes a whole subsystem's test surface, while the skill says tests protecting public APIs, plugin SDKs, protocols, storage, security, configuration, migrations and cross-language contracts should be retained. steipete said a cleanup agent needs a concrete target, offering "remove 20% of the least useful tests while maintaining code coverage within 2%" as the example.