Microsoft Brings Its Copilot Harness to Foundry

Microsoft is turning agent orchestration into shared infrastructure inside Foundry.

The AI Enabled PM

Oct 29, 2026

On October 28, Microsoft introduced its Copilot harness in Foundry Agent Service in public preview. It packages the agent loop, working context, sandboxed execution, skills, and connected tools as a managed service. Teams bring their instructions, procedures, and supported model choice.

Tweet

For product teams, this puts a build-or-buy decision inside the agent stack. Owning execution infrastructure takes engineering time that could go into domain workflows and evaluating outcomes. A managed harness is worth testing where its controls fit the product.

The preview has limits: existing agents cannot switch to the harness after creation, and Foundry manages CPU, memory, and idle settings. Separately, long-running resilience for hosted agents is generally available. Recovery still depends on customer code saving progress and preventing duplicate actions when work restarts.

Run one representative workflow against your existing stack. Compare accepted outcomes, operator intervention, recovery, and total cost. Include model and tool charges; harness runtime billing starts December 1. Keep your own loop where you need execution behavior the managed service cannot provide.

Sources

A Thousand Sandboxes. One Shared Memory.

Fresh agent runs can still inherit information through shared package services, artifacts and workspaces. Here is how to check the boundary.

A working app with a success check balanced on unstable foundations.

Your Vibe-Coded App Works. Is It Ready for Customers?

A field guide for deciding when a vibe-coded app can move from internal testing to real customers.

An agent-shaped key approaches a product’s access-controlled doorway.

Your Next User Is an AI Agent

Products built for people need new identity, authority, observability and recovery controls when AI agents become active users.

A message leads to a bounded yes-or-no decision, shown as a physical switch.

Jev Shows That Not Every AI Call Should Generate Text

Jev returns bounded choices and probabilities instead of prose. Its real lesson for product teams is where generation should end and ordinary software should take control.