update
Jun 2, 2026
By Teun
Microsoft launches Scout, an OpenClaw-inspired assistant
Microsoft has launched Scout, an AI assistant built on the OpenClaw framework for use inside Microsoft 365. The company says it is designed to learn from user feedback, run across desktop and web, and include security checks and audit trails.
Microsoft is launching Scout, a new AI assistant built on the OpenClaw framework and aimed at Microsoft 365 users. The company says Scout is designed as an always-on agent with a persistent identity, meant to work alongside a user and adapt over time based on feedback.
The product was announced in the first weeks of 2026 as part of Microsoft’s Build developer conference lineup, which also included Project Solara, an update to Copilot, and a new reasoning AI model. Scout is available through Microsoft’s Frontier program, which gives early adopters access to experimental Microsoft products, and it requires a GitHub Copilot subscription.
⚡ New to this?
Scout is Microsoft’s version of an AI assistant that can remember how a person works and do tasks over time, rather than answering one question at a time. OpenClaw is the framework Microsoft used to build it, and a policy conformance system is a safety layer that checks whether the agent is following rules.
This matters because AI agents are moving from chat tools to software that can act inside email, calendars, and other work systems. That raises both productivity gains and security questions, since a mistake can affect real accounts and real data.
🦞 OpenClaw angle
If you build self-hosted agents, treat memory and policy checks as first-class features, not add-ons. Give each agent a narrow identity, store user-specific preferences separately from task state, and log every policy decision so you can audit behavior later.
Also test agents against inbox, calendar, and browser workflows in a sandbox before connecting them to production systems. Microsoft’s approach suggests that persistent assistants will be judged as much on guardrails and traceability as on raw capability.
In Microsoft’s demo, each Scout instance can be given a name, such as “Sebastian.” According to Scout VP Omar Shahine, the idea is to let users encode the way they work into the agent’s memory and skills so it becomes more useful over time.
“We all have our interesting quirks in how we work, and people are codifying those patterns into memories and skills that persist in their agent,” Shahine said. “Then the agent becomes more capable, better understanding you and gaining more agency and exercising judgments.”
Microsoft says Scout is cloud-based, but it also works across the desktop and web browser. That design is meant to make it easier to connect the assistant to inboxes, calendars, and other systems that users already rely on.
The company is also shipping Scout with prepackaged skills, including calendar management and drafting meeting agendas. Shahine said Microsoft expects the bigger value to come from the skills users create themselves, rather than only the built-in features.
That learning loop is central to the product. The more a user trains the assistant with repeated behavior and feedback, the more useful Microsoft says it should become. The company is betting that this kind of personalization will make the assistant harder to replace once it has learned a user’s routines.
Microsoft is also emphasizing security. The company said Scout includes a built-in “policy conformance system” that continuously checks whether the agent is operating according to set guidelines. Each conformance check creates its own audit trail, giving users and administrators a record of what the assistant did and whether it stayed within policy.
Those controls appear aimed at a real concern for agentic systems: software that can act without constant supervision. The source article points to earlier OpenClaw-related examples where an agent behaved erratically inside a researcher’s inbox, underscoring why Microsoft is highlighting guardrails.
Scout is the latest example of Microsoft folding OpenClaw’s agent model into its own product line. For now, the assistant is being positioned as an experimental tool for early adopters, with Microsoft using Frontier to test how far a personalized, persistent AI assistant can go inside its productivity stack.
