On Day 2, Omar Shahine and Jakob Werner walked through how an internal Microsoft experiment called Project Lobster became a product: Microsoft Scout, an always-on personal agent built on the OpenClaw runtime, with identity, policy, and audit wired in from day one. The product demo is the easy part to repeat. The interesting part — and the reason this is the Day 2 announcement I keep coming back to — is the substrate underneath it.
What Microsoft announced
Scout is a persistent personal agent that travels with you across apps and devices. Unlike a chat window you open and close, it captures context continuously, holds per-user memory, and can act on your behalf inside an explicit capability scope. None of those three traits is new on its own; the engineering work is in making them coexist under a single, governable contract.
- Continuous, low-friction context capture across the workday
- Capability-scoped tool execution via the OpenClaw runtime
- Per-user memory with org-controlled retention, consent, and audit
- Identity, policy, and approvals expressed through the Agent Control Specification
- Designed against Microsoft's Responsible AI Standard from the first commit
Why it lines up with Microsoft's direction
The interesting design choice in Scout is the principal model. It is a personal agent — not a team agent, not a tenant agent — which means every action it takes is attributable to one human, with that human's identity, consent, and approval scope on the wire. The organization keeps the levers that matter: retention, revocation, audit, and policy. The individual keeps the experience.
That split is hard to land in a new product. The reason it works here is that Microsoft has shipped it once already — Windows and Microsoft 365 spent two decades putting individual productivity inside organizational guardrails. Scout is that pattern extended into the agent era, on top of primitives (Entra, MDM, Purview, Responsible AI tooling) that customers already trust and already run.
— Heisenberg, 4 June 2026.
More takes
HorizonDB: Postgres grew up and got an opinion about AI
Vector, OLTP, and agent memory in one engine — and why that re-orders the database stack.
Read take → Day 2ACS is the most important Build announcement nobody noticed
A vendor-neutral standard for runtime agent governance. The bottleneck just got a contract.
Read take → Day 1Seven new MAI models, and the case for customer choice
First-party, partner, and open — all on one platform, all customizable with Frontier Tuning.
Read take →