Agents are not new. Microsoft has done extensive research in this area and last year created a multi-agent library for developers around the world. This has helped shape the capabilities agents can perform today. Recent advances in large-scale language models (LLMs) have brought even more attention to AI by allowing anyone outside the developer community to communicate with it. This combination of agents and LLM makes AI tools more specifically useful.
“People expect AI to do things for them, not just generate language,” said Eas Comer, managing director of Microsoft’s AI Frontiers Lab. “If you want to have a system that can actually solve real-world problems and help people, that system has a good understanding of the world we live in, and when something happens We must recognize the changes and act accordingly.”
Agents are like a layer on top of a language model, observing and collecting information, providing input to the model, and together generating and communicating an action plan to the user. Alternatively, they may act alone if permitted. Therefore, both agents and models are equally important pieces of the puzzle as far as generative AI tools are concerned.
Agents will become more useful and have more autonomy through innovations in three necessary elements: memory, qualifications, and tools.
Memory provides continuity, so every time you ask for something, it’s different from starting from scratch.
“To be autonomous, context needs to be conveyed through a sequence of actions, but the model is very disconnected and doesn’t have continuity like we do. That’s why all the prompts “It creates a vacuum and can trigger false memories,” he says. said Sam Schillace, Microsoft Deputy Chief Technology Officer. “It’s like watching stop-motion animation frame by frame, and it starts moving in your head. Clay models don’t move on their own.”
To build a memory infrastructure that addresses this, Schillace and his team are working on a process of chunking and chaining. This is basically what it feels like. They are experimenting with breaking down interactions into bits that can be stored and linked by association, similar to memory, for faster access. For example, group conversations about a particular project so that agents can remember the details at any time. Request a status update so you don’t have to search the entire database.
Credentials and tools allow agents to securely access information (for example, who your boss is) and computer programs they need to get things done for you, with your permission. You will be able to access or access that information securely. Something like Teams or PowerPoint needs to perform an action on your behalf.
How to use and build a work agent
Microsoft 365 Copilot makes creating and publishing agents that help you do your daily work as easy as creating a spreadsheet or presentation. No coding skills required.
You also don’t need to be a developer to build agents using Copilot Studio. Anyone can connect to relevant business data such as emails, reports, and customer management systems so they can perform tasks and gain insights.
You’ll also be able to register new agents with Microsoft 365 to help with common workflows and tasks. Interpreter in Teams provides real-time speech translation during a meeting, for example, and you can choose to simulate your own voice. Employee self-service agents simplify HR and IT help desk-related tasks, such as helping employees troubleshoot problems with their laptops or making sure they’ve maxed out certain benefits. Masu. You can also connect to your corporate systems for further customization with Copilot Studio.
Microsoft Dynamics 365 also includes agents responsible for a variety of common business workflows across sales, supply chain, finance, and customer service functions.
And soon, every SharePoint site will have an agent tailored to your organization’s content. This allows employees to quickly tap into these vast knowledge bases and find exactly what they need in seconds, whether it’s project details buried in a workback schedule or a project overview. You will be able to do this. Recent product notes.
Developers have even more options. With the new Azure AI Agent service, choose from small or large language models to tune, develop, and extend agent-powered apps to streamline complex workflows like order processing and customer data syncing. and can be automated. We provide a software development kit with tools to develop agents and efficiently integrate agent functionality using Visual Studio Code and GitHub.
One of the models, OpenAI’s recently announced o1 series, provides agents with more advanced reasoning capabilities to help IT help desk personnel complete tasks such as obtaining the information they need to solve a problem. Allows you to take on more complex tasks by breaking them down into steps. Solve the problem and make a plan taking into account the solutions you have tried.
You can also harness the power of LinkedIn agents. The platform’s first agent helps recruiters with hiring.
Risk assessment for autonomous action
Agents that can act autonomously require special safety considerations, and Microsoft is focused on ensuring agents only have access to what users want, the company said in its Responsible AI. said Sarah Bird, Chief Product Officer.
“The stakes are certainly higher for agents from a responsible AI perspective,” Bird says. “So the error rate needs to be much lower. Plus, there are a lot of nuanced situations where something could be an error. This is a big challenge for agents.”
But other AI applications can use a similar Responsible AI Fundamentals playbook to assess and mitigate risk for agents, she says.
The new Copilot Control System helps IT departments manage Copilot and agents with data access and governance, management and security controls, and measurement reports and tools to track adoption and business value.
Many agents, such as those created for Microsoft 365 and Dynamics 365, include “human” approval, such as those responsible for the final step of reviewing and sending an email created by a sales order agent. must be performed by a person. Additionally, for agents developed in Copilot Studio, authors can review records to see what actions the agent took and why.
The key is for organizations to choose the right starting point for their needs, with a focus on testing and moderation to ensure accuracy, Bird says.
“Of course, we progress by building on the foundations we already have, so we start our journey from a strong point,” Bird says.
Looking back at the past and looking towards the future
Comer, who has been developing AI agents since 2005 and wrote his Ph.D., said engineers have long been excited by the idea of autonomous systems working with and assisting people. The hurdle, she says, was that the backend “lacked general problem-solving capabilities.”
The LLM “finally gave me the missing component,” she says. “Now we can bring back many ideas from decades of research.”
Going forward, Kamar envisions a new ecosystem or marketplace of agents where apps will allow people to do more with their smartphones.
Agents already have “the basic building blocks of what they need to complete a task,” she says. “It’s like observing, ‘Looks like the meeting is taking too long. We should delay the next meeting.’
As they gained autonomy through innovations in memory and rights, they became even more useful. We ease employee headaches by assisting with expense reporting, project management, meeting facilitation, and more. They are also having a dramatic impact on businesses by alerting supply chain managers to inventory shortages and automatically reordering to increase sales and keep customers happy.
The key to agents is that they “open up a set of opportunities to collaborate with people to complete tasks, and that’s what we expect from AI systems,” Comer said. “AI agents will not only be a way to bring more value to people, they will be a paradigm shift in how work gets done.”
And this is just the beginning. Copilot will evolve with new features such as Copilot Actions. Copilot Actions is designed to handle the mundane tasks that get employees stuck, like summarizing emails they missed while on vacation, editing agenda items, and writing monthly reports. Over the next year, we’ll be adding more features like this to make work easier for your employees and teams.
“Copilot enables every employee to do their best work in less time and focus on more meaningful tasks,” Spataro said. “And agents created with Copilot Studio can transform any business process, helping companies streamline operations, enhance collaboration, and drive innovation at scale.”
Illustration: Michał Bednarski / Makeshift Studios
