AI alone won’t change your business. The system running it will.

AI is making a significant mark in the business world, leading to rapid changes across all functions and roles. Along with this transformation, a new type of organisation is emerging—one that looks markedly different from the traditional companies we’ve known. The key to success won’t lie in merely showcasing demos but in effectively integrating AI into a controlled, continuously evolving system that enhances real-world operations.

This evolution extends far beyond just utilising chatbots. While these tools can be helpful, they don’t fundamentally change the operations of large organisations. The real potential lies in organising teams of AI agents that can seamlessly work across various functions, including software delivery, customer support, finance, human resources, and operations, while ensuring that there’s enough identity, context, policies, and human oversight to foster trust in their output.

For this vision to succeed, businesses need more than just top-notch AI models or scalable computing facilities. The success of AI hinges on the entire system surrounding it: from how agents are constructed and deployed by engineering teams to their integration within the company’s context, a robust governance structure, and mechanisms for continuous, safe improvement over time. Without a cohesive system, AI becomes scattered, fragile, and hard to rely on at scale.

Our approach is fundamentally different. We’re creating a holistic agent platform that supports various models and prioritises flexibility and choice at every layer. Central to this design is the developer experience. Today, we’re excited to share that significant advancements in this platform are coming together.

Building a System for the Enterprise of the Future

To thrive in this new landscape, an agent platform must exceed traditional limitations. It needs to handle genuine production workloads, navigate complex organisational structures, and take on real business responsibilities.

We’re focusing on three essential principles:

First, it should be a unified, integrated system that accommodates a variety of models.
Organisations can’t afford to piece together their AI strategy in a haphazard way. Using disconnected tools that are integrated retroactively can slow teams and increase risk. Therefore, all aspects—building, context integration, operation, governance, and improvement of agents—must occur within a unified system. That’s why we’re merging Azure, GitHub, Microsoft IQ, Fabric, Foundry, Windows, Microsoft Security, and Microsoft 365 into a single platform, ideally suited for scaling agent deployment. It’s essential for companies to choose the right model based on the task at hand, balancing quality, speed, and cost—utilising Microsoft, partner, and open models.

Second, it must be secure and governed by design.
While it’s easy to talk about governance, delivering it is much more complex. Achieving real governance means establishing a coherent stack that spans from development through to production, built on trusted foundations of identity, access, compliance, and security. By integrating Entra, Purview, Defender, Agent 365, and the broader Microsoft Security suite, we ensure that governance is inherent to the system, rather than an afterthought, allowing an AI-first enterprise to flourish without compromising control.

Third, it must be designed for continuous enhancement.
AI systems in enterprises can’t remain stagnant. The behaviours, outcomes, and feedback from human users must feed back into the system, allowing it to evolve safely over time. As the system operates, its models, workflows, and agents grow more sophisticated and tailored to the specific processes of the business. This growth results in increasing value as time goes on.

These features are becoming essential, and companies that align their AI strategies with these three principles will gain a significant advantage, achieving results in months rather than years.

But what does a system like this look like in action? It all begins with how agents are constructed. Let’s explore that process within our platform.

1. Build with GitHub

GitHub is where your developers are already working. It’s the hub for all your dependencies, where your application code resides, and where collaboration with the open-source community happens. Developing agents elsewhere would mean abandoning these crucial resources.

Agents should be constructed just like production software. Use GitHub Copilot to speed up coding. Gather all essential resources: codebases, work items, agent skills, and tools. Remember, agents aren’t just code—they also require evaluations and monitoring tools, all versioned systematically like any production-grade system.

Agents must follow a lifecycle: source, test, deploy, observe, and improve. GitHub sets this lifecycle in motion, offering the necessary controls right from the start. The outcome is a streamlined workflow for agent creation, complete with proper guardrails at all times, all within a new app tailored for this purpose.

2. Contextualise with Microsoft IQ

Code is just one part of an agent’s makeup. For it to truly be efficient, the agent must understand your business nuances: customers, products, contracts, and processes. Without reliable enterprise context and intelligence, even the most advanced model can only make educated guesses.

Businesses require a wide array of models and the ability to assign the right model to the right job. However, simply having options isn’t enough. Microsoft IQ grounds agents within the context of your enterprise data, connecting to your core business systems and resources. This includes accessing data across Microsoft 365, customer data, knowledge bases, and even websites. With Web IQ, the newest addition to the IQ platform, agents can also pull relevant online information when necessary.

But it’s not just about access; directing AI toward unprocessed data is inefficient. Microsoft IQ organises and secures information, presenting it to agents in a usable format, helping them gain accurate insights without being overwhelmed by irrelevant data.

Once agents are established within the correct context, businesses can take it a step further. With Frontier Tuning, you can fine-tune how AI models behave using your own data and workflows.

This includes Microsoft’s seven new MAI models, which cover a range of tasks like imaging, voice recognition, transcription, coding, and reasoning. Importantly, these models are adaptable and designed to learn from actual business operations.

Our reinforcement learning environments act as training grounds for AI. Here, agents learn your specific processes and standards, allowing for a custom fit that enhances your return on investment (ROI).

Moreover, your custom or post-trained models remain within your environment, meaning your intellectual property and proprietary data inform how your agents operate. The resulting intelligence remains under your control, with proprietary learning processes integrated directly.

Without this context and Frontier Tuning, agents are generalists. When equipped with these features, they transform into specialised partners deeply familiar with the enterprise landscape.

3. Run with Foundry

Once agents are built and contextualised, they require a reliable environment for execution—not just as an experiment, but at scale.

Agents impose different demands on a runtime compared to traditional applications. They need to reason, perform actions, call upon tools, coordinate with one another, and adapt over time—all while adhering to enterprise guidelines. Foundry is tailored for this reality.

• A diverse range of models: Different agents have distinct strengths, price points, and tasks. Foundry gives access to the necessary models and an optimised routing system to balance quality, speed, and costs effectively.
• Enhanced performance for open models: Leveraging Fireworks AI on Foundry, enterprises receive faster, more efficient inference directly within the platform.
• Support for agents outside our framework: Integrate agents developed through the Microsoft Agent Framework, LangGraph, GitHub Copilot SDK, Claude Agent SDK, or a custom harness.
• Tools and actions: Agents interact with enterprise systems using MCP, connectors, APIs, and workflows, ensuring safe execution by default.
• Evaluation and tracing: Monitoring tools and traces allow for measurable agent performance. If you can’t measure it, you can’t improve it.
• Continuous optimisation: Foundry supports the refinement of models, tools, IQs, and actions over time, enhancing agent performance consistently.

A comprehensive trust, security, and policy framework envelops the entire runtime. Policies are uniformly applied regarding context access, tool usage, optimisation updates, and response delivery. The agent carries out its tasks exactly as your enterprise requires.

This is the crossover where your agent transitions from a mere project to a robust production system.

4. Govern with Agent 365

Now imagine scaling that single agent to hundreds or even thousands. This is the scenario as various teams across a business develop their agents. Some may be well-crafted, while others might lack structure, provide excess access, or operate independently without contributing to broader goals.

Effective governance isn’t optional in large enterprises. Businesses require a way to keep tabs on what’s running, what access it has, and how tasks are being performed, enforcing policies across all agents.

Agent 365, combined with Entra, Purview, Defender, and the Microsoft Security Suite, provides this capability. If security is a priority alongside AI, consider exploring “MDASH.”

Every agent deployed across the organisation can be found in a single catalogue, regardless of whether it was created in Foundry or elsewhere. IT teams gain visibility into who deployed each agent, which data and tools it can access, and its performance metrics, enabling quick corrective actions when necessary.

One centralised location for complete visibility and control over your agents’ actions.

5. Commit to Continuous Improvement

Agents must not be static entities. Each agent’s action generates useful data: trajectories, outcomes, and feedback. The system collects, refines, and feeds these insights back into the improvement cycle: observe, evaluate, enhance, and safely deploy changes.

This learning process runs continuously in real-time.

Most gains stem from improving agent functionality through evaluations of prompts, context, skills, and tools. As patterns are identified, learning can progress into model routing across multiple frameworks, fine-tuning parameters, or reinforcement learning, always anchored in an evaluation process to boost quality and ROI as required.

This loop is governed, not closed—enterprises need the ability to audit, adjust, and control how changes are implemented. The system will grow more proficient over time, directed by human oversight while becoming more autonomous, yet never out of reach.

This method represents an ongoing improvement strategy, where the overarching system evolves in real-time.

6. Integrate within the Workflows of Your Team, Scalable on Azure

Ultimately, none of this progress matters if it does not reach the people who carry out the work.

Agents show up directly within the daily workflow, in Teams, across Microsoft 365, and within your own applications and experiences. Security, identity, and compliance are built in from the ground up, ensuring that the agents your teams depend on inherit the same trustworthy framework as the rest of your environment.

Multiple platforms are supported, but your agents can be optimally developed and run securely on Windows. You have the option to execute models either in the cloud or locally, with top-tier sandboxing safeguarding always-on agents.

When you need AI-optimised computing, a reliable global infrastructure, or a pathway to market, this system scales effortlessly on Azure—the same trusted foundation that customers have relied on for years.

Compounded Value of the System

All forward-thinking enterprises will converge on this approach: a centralised AI platform that coordinates activities across the business, merging data, models, agents, and human judgement into a continually improving, secure system.

As this system operates, its value grows exponentially. Efficiency increases, shifting the constraints from manual effort to human creativity and collaboration. Employees can accomplish more independently, guided by shared context and fewer arbitrary handoffs, allowing the business to accelerate while adding minimal friction.

We’re experiencing a period of considerable upheaval. The enterprises that thrive in this evolving landscape will be the ones that dynamically adjust to changing conditions, simplify processes across the business, and effectively transform intelligence into tangible outcomes. Microsoft’s agent platform is crafted to achieve just that, facilitating the building, contextualisation, executing, governance, and ongoing enhancement of agents as a unified system.

At this stage, the platform transcends being merely a construction layer; it evolves into the foundational operating system for scalable enterprise AI, where intelligence and trust are inherently embedded.

Tags: AI, GitHub, Microsoft Agent 365, Microsoft Build