11/17/2026, 07.52
by Cristinacovers work, organisations and regulation
Condividi su Facebook Condividi su Twitter Condividi su Pinterest Condividi su Telegram Condividi su WhatsApp

Microsoft Opens Copilot Harness in Foundry Agent Service for AI Agents

Microsoft launches the Copilot harness in Foundry Agent Service, providing a managed orchestration layer to build complex, multi-model AI agents for enterprises.
Microsoft Opens Copilot Harness in Foundry Agent Service for AI Agents
Key points
  • Microsoft introduces the Copilot harness in Foundry Agent Service (public preview), offering a managed execution layer for AI agents.
  • The system decouples the AI model from the orchestration, supporting over 20 frontier models and BYOK configurations.
  • New capabilities include sandboxed code execution, subagent delegation, and general availability of long-running resilience.
  • Developers can now deploy agents that handle complex, multi-step workflows without building the underlying agent loop from scratch.

The transition from a simple AI chatbot to a functional AI agent requires more than just a powerful Large Language Model (LLM). While a model provides the reasoning capability, the actual execution of a task demands a surrounding infrastructure: a way to manage memory, call tools, handle errors, and maintain state across long durations. Recognizing this gap, Microsoft has announced the public preview of the Microsoft Copilot harness in Foundry Agent Service.

This move effectively democratizes the same orchestration engine that powers Microsoft's most successful AI products, including Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, and Visual Studio Code. By exposing this harness through Microsoft Foundry, the company is shifting the developer's burden from building the plumbing of AI agents to designing their actual utility.

The architecture of a managed agent harness

In technical terms, an agent harness is the runtime environment that transforms a static model into an active agent. Most development teams currently spend significant resources implementing repetitive execution machinery: coordinating model and tool calls, managing working context, and deciding how to proceed when a task requires multiple sequential steps. The Copilot harness absorbs this complexity.

Under this new model, Microsoft manages the agent loop and the sandbox execution, including governance and optimization. The developer, meanwhile, remains in control of the high-level logic. This means the user provides the instructions, the specific skills, the tools, and the choice of model, while the harness ensures those elements work in harmony to complete a goal.

The system is designed to be declarative. Instead of writing exhaustive scripts for every possible outcome, developers define agents that can break a goal into smaller steps, access knowledge via connectors, and adjust their trajectory if a specific step fails or the user's request changes mid-process.

Breaking the model lock-in cycle

One of the most persistent pain points in AI development is the orchestration rewrite. Traditionally, switching from one model to another often requires a complete overhaul of the orchestration layer because different models handle tool calls and context differently. The Copilot harness aims to eliminate this friction by acting as a universal adapter.

The harness supports more than 20 frontier models across various families, including the ability to use Bring-Your-Own-Key (BYOK) configurations. This multi-model routing capability allows the system to direct specific tasks to the most appropriate model based on the complexity or nature of the work. For instance, a developer might deploy a high-reasoning model like gpt-5 for complex planning and a smaller, faster model like gpt-5-mini for routine data retrieval, all within the same agent workflow without rewriting the underlying code.

Core capabilities and the sandbox environment

The public preview introduces four primary building blocks that expand what an agent can actually do. First, reusable skills allow developers to package instructions and scripts into procedures that agents can apply consistently. Second, connected tools integrate both Foundry-supported tools and those following the Model Context Protocol (MCP), ensuring that authentication and approvals are explicitly configured.

The third and perhaps most critical component is execution and delegation. The harness provides a sandboxed session where agents can read and write files in the agent workspace, perform web searches, and run shell commands. Furthermore, the main agent can spawn and delegate bounded subtasks to built-in subagents, coordinating the final result without losing the primary thread of the conversation.

Finally, agent definitions are source-controlled, meaning they can be managed and versioned alongside standard application code, bringing AI agent development closer to traditional software engineering best practices.

Solving the problem of long-running workflows

A major limitation of early AI agents was their fragility during long-running tasks. If a process was interrupted or a network error occurred, the agent would often lose its place, forcing the user to restart the entire sequence. Along with the harness preview, Microsoft has announced the general availability of long-running resilience for hosted agents.

This feature allows agents to save their progress and intermediate results. If a workflow is interrupted, the agent can resume from where it left off rather than starting from scratch. This is a critical requirement for enterprise-grade automation, where tasks might span hours or require asynchronous approvals from human operators.

The Microsoft Copilot harness is the managed orchestration layer around the model that turns reasoning into completed work. If the model supplies the intelligence, the harness is everything around the model.

Implementation paths for developers

For those looking to integrate this into their stack, Microsoft has provided multiple entry points. Developers can author agents via the Foundry portal, the Foundry Tool, or the preview Azure Developer CLI (azd). The transition is designed to be flexible; the Copilot SDK is available for download, and while the harness can run in the cloud managed by Foundry, Microsoft has stated that its platforms remain open to other harnesses and frameworks.

However, early adopters should be mindful of the rollout timeline. While the runtime is free until December 1, 2026, the system is currently in public preview. This suggests that while it is an ideal environment for experimentation and prototyping, enterprises may want to maintain their current production stacks until the preview limits are fully resolved.

Strategic implications for global enterprises

For businesses in the USA, UK, and other global markets, the introduction of the Copilot harness signals a shift toward the industrialization of AI. We are moving away from the era of the standalone prompt and into the era of the autonomous agentic workflow.

In the US and UK markets, where the focus is heavily on operational efficiency and reducing technical debt, the ability to swap models without rewriting orchestration is a significant competitive advantage. It prevents vendor lock-in and allows companies to pivot as new, more efficient models emerge from the frontier. From a regulatory perspective, the inclusion of built-in governance, approvals, and sandboxed execution addresses key concerns regarding AI safety and data leakage, which are paramount for firms operating under strict financial or healthcare compliance frameworks.

By providing a standardized execution layer, Microsoft is essentially creating a blueprint for how enterprise AI should be deployed: intelligence provided by the model, but control, safety, and execution managed by a rigorous, version-controlled harness. For the global entrepreneur, this means the barrier to deploying complex AI agents has dropped significantly, shifting the value proposition from who has the best engineers to build the loop, to who has the best domain expertise to define the skills.

FAQ

What is the difference between a model and a harness?

The model provides the reasoning and intelligence (the brain), while the harness provides the execution environment, tool coordination, and memory management (the body and nervous system) required to complete a task.

Does the Copilot harness support models other than OpenAI?

Yes, it supports more than 20 frontier models across different families and allows for Bring-Your-Own-Key (BYOK) configurations.

What is long-running resilience?

It is a feature that allows hosted agents to save their progress and intermediate results, enabling them to resume interrupted work instead of restarting the entire process.

Is the Copilot harness currently free to use?

The runtime is free until December 1, 2026, as part of the public preview in Microsoft Foundry.


Sources: Devblogs, Michaelscollier, Aireiter ·

Hai una domanda su questo dossier?

Scrivila qui: Susanna, l assistente AI di glacom, ti risponde via email con un approfondimento gratuito.

Nessuna consulenza personalizzata (finanziaria, legale o medica): solo informazione e fonti. Email usata solo per rispondere.

oppure scrivile su: WhatsApp · Telegram · SimpleX · Delta Chat · Email

Condividi su Facebook Condividi su Twitter Condividi su Pinterest Condividi su Telegram Condividi su WhatsApp
Printable version
CLOSE X
See also
Anthropic's Claude AI Accelerates Drug Development Pipelines
Anthropic reveals how Claude AI is slashing drug batch review time by 89% and designing thousands of antibodies through partnerships with six biotech …
17/11/2026 13:19
GitHub Copilot Fusion: Slashing AI Coding Costs via Orchestration
GitHub debuts Copilot Fusion, an orchestration engine that mixes OpenAI and Anthropic models to match top performance at nearly half the per-task cost…
17/11/2026 11:06
Google Gemini 4 Carbon: New Frontier Model Claims AI Top Spot
Google unveils Gemini 4 Carbon, targeting cybersecurity and long-horizon engineering. It leads the Artificial Analysis Index but remains in limited re…
17/11/2026 09:50
Bitly Data: Short Links Aren't the Main Cause of Conversion Loss
New Bitly research reveals that while generic short links can cause hesitation, the real conversion killer is directing social traffic to generic home…
16/11/2026 19:05
WordPress 7.1.4 Security Update: Fixing RCE and Core Flaws
WordPress releases version 7.1.4 to patch 22 vulnerabilities, including a critical remote code execution risk in the update process. Immediate update …
16/11/2026 09:57