New Microsoft Copilot Harness | More power, token efficiency & control for agents in Foundry

Build agents you can trust with your company's data, apps, and resources using Microsoft Foundry. Start with a fully managed prompt agent and move it onto the Microsoft Copilot harness so it plans, loops, calls subagents, and delivers finished reports. You can also bring your own harness, such as the Claude Agent SDK, and run it as a hosted agent with isolated VM sandboxes, autoscaling, and zero idle cost.

Give every agent its own Entra Agent ID, restrict its egress to approved endpoints, and trace every action it takes. Curate tools once in a governed, MCP-compatible toolbox, cut token use with tool search, and tune agents against your own benchmark with Agent Optimizer.

Jeff Hollan, Microsoft Foundry VP, shares how to balance agent autonomy with enterprise control from first build through production.

👥 Who it's for: 
Developers and AI engineers building production agents, solution architects designing agent platforms, and IT and security teams responsible for governing agent identity, access, and network egress.

⏱️ Chapters: 
00:00 Why the harness decides agent trust
00:59 What an agent harness does
01:22 Five components built into every harness
02:15 Microsoft Copilot harness meets Foundry controls
03:28 Where a prompt agent falls short
04:57 Same agent on the Microsoft Copilot harness
06:10 Guardrails block unauthorized agent actions
07:29 Entra Agent ID and egress rules
08:05 Bring your own harness as a hosted agent
09:06 Private networking, regions, and your own registry 
10:13 Curate tools once with Foundry toolboxes
11:13 Tool search keeps agent context lean
11:54 Benchmark agents against your real tasks
12:54 Auto-tune agents with Agent Optimizer
13:53 Start building in Microsoft Foundry

What is an agent harness? The harness is the environment you wrap around your model and agents. It runs the agent loop, manages context as it grows, provides skills and tools such as MCP servers, delegates work to subagents in parallel, and maintains memory across turns and sessions. The Microsoft Copilot harness is model-agnostic, optimized for enterprise task completion, and runs behind Microsoft services like GitHub and Microsoft 365.

Inside Microsoft Foundry, it respects your existing permissions and data protection controls and runs each agent instance in its own isolated sandbox. Hosted agents in Microsoft Foundry Agent Service also support Hermes, OpenClaw, and fully custom harnesses, along with your own virtual network, container images from your own registry, and expanded regions so agents run where your data is.

🔗 Related links:

► Head to Microsoft Foundry to build your first project at https://ai.azure.com

►For hosted agent templates and the SDKs to start coding, check out https://github.com/foundry-samples

► Unfamiliar with Microsoft Mechanics?
Microsoft's Official Video Series for IT

— Subscribe https://www.youtube.com/c/MicrosoftMechanicsSeries
— Microsoft Tech Community: https://techcommunity.microsoft.com/t5/microsoft-mechanics-blog/bg-p/MicrosoftMechanicsBlog
— Podcast: https://microsoftmechanics.libsyn.com/podcast

►Join us on social:

— https://twitter.com/MSFTMechanics
— https://www.linkedin.com/company/microsoft-mechanics/
— https://www.instagram.com/msftmechanics/
— https://www.tiktok.com/@msftmechanics