This post was originally published on this site.
Microsoft on Wednesday introduced out two “reasoning agents” it claims can handle research and analysis projects.
In a departure from Redmond’s convoluted naming conventions, they’re identified by their function, Researcher and Analyst.
Much like the 11 Security Copilot agents introduced earlier this week, Researcher and Analyst can make decent predictions and automate multi-step tasks that utilize the often adequate predictive competency of machine learning models and whatever data sources have been surrendered to the AI.
“Recent advances in LLMs allow AI to contribute like a team member,” said Ashok Kuppusamy, corporate VP for M365 Core Experiences, in an explanatory video. “We’ve taken one of these models – OpenAI’s o3 deep research – along with advanced orchestration in Copilot and deep search algorithms and optimized it for work.”
If your job involves development of product strategies, or writing reports about business operations, take note because Researcher can compile such documents.
The Register is unaware of any company that has achieved improved results using an AI-authored business plan, though presumably that’s the goal. At least if Researcher’s sample advice [PDF] to fictional wearables biz Proseware flops in the market, a human scapegoat may not be necessary.
Researcher, according to Kuppusamy, takes a prompt and then tries to answer it using a combination of OpenAI’s deep research model, Microsoft Graph data, and external sources including the web, and data stored in applications including Salesforce, ServiceNow, and Confluence.
Analyst is oriented toward data analysis. Kuppusamy in a video showed off how it might help visualize an Excel spreadsheet of customer data.
“Usually, to make sense of this data I’d need to ask my colleague who knows Python,” he explained. “But let’s try the Analyst agent. I don’t have to spend time writing the perfect prompt to get what I’m looking for. I’ll just ask it to help me come up with an easy way to learn and visualize my customer base.”
Through its chain-of-thought process, the Analyst agent will work through the prompt in a series of steps and even generate Python code to create the requested graph.
“I can always click and see its chain-of-thought reasoning, as well as the Python code it’s running in real time,” Kuppusamy continued. “This allows me to validate and trust in its thinking and approach.”
The two agents will be available next month through Microsoft’s new “Frontier” program that makes Copilot enhancements available to customers while they’re still in development.
Meanwhile, Microsoft has expanded the capabilities of its Copilot Studio agent-building tool by adding deep reasoning and agent flows.
Deep reasoning enables CoPilot’s underlying AI model to break down complex instructions into a sequence of steps and evaluate each one in the context of the task’s overall goal. This step-by-step reasoning takes more processing time but can deliver more accurate and context-aware results.
Agent flows allows agents to follow predefined logic paths, helping ensure interactions stay on track and aligned with business rules. For instance, an AI agent handling customer feedback could be configured to apply specific escalation policies and trigger a human response for high-priority issues.
Finally, autonomous agents are now available within Copilot Studio. These are self-directed bots that can plan tasks, adapt to changes, and escalate issues to humans without waiting for a prompt. ®