Microsoft Copilot harness manages the agent loop in Foundry
Microsoft has put its managed agent harness into public preview in Foundry Agent Service, with reusable skills, connected tools and consumption billing from December 1, 2026.
On October 28, 2026, Microsoft introduced the Microsoft Copilot harness in Foundry Agent Service as a public preview, and said long-running resilience for hosted agents is generally available this week. The post on the Microsoft Foundry blog presents the harness as a fully managed way to build agents that reason, call tools and finish long, multi-step tasks.
What sits inside the harness
Microsoft lists four building blocks. Reusable skills package instructions, scripts and supporting resources into procedures an agent can apply across tasks. Connected tools cover supported Foundry tools and Model Context Protocol (MCP) tools, with authentication and approvals configured explicitly. Execution and delegation let an agent work with files and code, and hand bounded pieces of work to built-in subagents while the main agent coordinates the result. Source-controlled agent definitions can live in Foundry or sit alongside application code.
Microsoft runs the agent loop and the sandbox; the team brings knowledge, procedures and tools. The harness holds working context and sequences model and tool calls inside configured limits and policy boundaries, and a supported Foundry model can be changed without rewriting the orchestration around it.
What changes for a team building agents
Much of this is undifferentiated work: coordinating model and tool calls, holding context, managing files, deciding what happens after step one. Microsoft's argument is that teams should put that time into the agent's expertise and its evaluation instead. The post quotes Confluence Technologies CTO Chris Evans saying a solution was running in less than a week, and Fashable CEO Orlando Ribas Fernandes describing governed sourcing workflows that keep commercially binding decisions with humans. These are customer statements in the announcement, not independent results.
A research brief, defined in files
The post walks through an agent that researches a topic and produces an evidence-backed brief, using the preview Azure Developer CLI (azd) surface. A skill folder holds a SKILL.md file with the method: fix the question, audience, scope and time period; gather evidence through approved tools and cite sources and dates; separate facts from interpretation and flag conflicting evidence; use Code Interpreter for calculations or charts; return findings, evidence, limitations and open questions. It also tells the agent to treat retrieved content as data rather than instructions, and to keep confidential information out of web queries.
An azure.yaml file declares the skill and the agent, naming the harness type, a model supplied through an environment variable, and web search and Code Interpreter tools. Deployment is azd deploy --all, then azd ai agent invoke. The Foundry portal, Foundry Toolkit for Visual Studio Code and a Foundry plugin inside the GitHub Copilot app are listed as alternative entry points.
Preview limits and pricing
The harness is selected when an agent is created. Existing agents are unchanged, switching harness after creation is not supported yet, and Foundry manages CPU, memory and idle settings for the execution environment. Billing begins on December 1, 2026 on a consumption basis. Harness runtime usage is free before then, while model inference and separately billed tools and connected services keep their own charges. From December 1, sandbox compute is USD $0.0994 per vCPU-hour and memory is $0.0118 per GiB-hour, and prices may vary by region and agreement.
Hosted agents: recovery is partly your job
For hosted agents, where you bring the framework, agent loop, middleware and runtime behaviour, Foundry detects interruptions and restarts the work with its original input. Your code has to save progress and intermediate results so the agent can resume, and any step that changes another system needs a guard against repeating on restart. The post notes that conversation history alone will not restore work in progress, and gives the example of an agent saving completed stages of a dataset analysis and checking whether a report already exists before creating it again.
What it means for building apps with AI
Microsoft frames the choice as three levels: prompt agents without the harness, prompt agents with it, and hosted agents. The harness takes away infrastructure work. It does not decide what the agent should remember, which tools need approval, or how a restart is made safe.
If the AI you are building is a feature inside a web app rather than an agent platform, that split still applies. Draftship is an AI app builder: you describe an app, it writes and hosts the code, with a built-in database and sign-in. Its documentation on adding AI to your app covers that path. Draftship is not affiliated with Microsoft.
Primary sources
- Introducing the Microsoft Copilot harness in Foundry Agent Service · Microsoft Foundry Blog, Oct 28, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.