What events in the field of AI took place during the week

Global developers have introduced artificial intelligence (AI) agents capable of autonomously performing work tasks, conducting in-depth research and creating medicines. All the details are in the Izvestia article.

OpenAI has released new agents inside ChatGPT

OpenAI has launched "work agents", a new type of tools inside ChatGPT for corporate teams. Unlike previous versions, these assistants work independently in the cloud, maintain context between sessions, and continue to perform tasks even when the user is offline.

Agents get access to the code, their own file repositories, and integration with third-party services. They can also independently review software applications, collect customer reviews, prepare weekly reports, and analyze notes from negotiations to prepare draft letters.

Google has developed two agents for research

Google has introduced two research agents, Deep Research and Deep Research Max, based on the Gemini 3.1 Pro model. The regular version is optimized for quick response, while the Max version is designed for deep background analytics, such as preparing a detailed market report while the user is asleep.

Goals and objectives: agencies have established standards for the implementation of neural networks

Why has the bar been set low for business and how will efficiency be determined?

An important update was the support of the model context protocol, which allows agents to connect to closed databases and financial feeds. The developers claim that in deep mode, their product surpasses its competitors in terms of the quality of information search on the Web.

In addition, agents can generate graphs and infographics inside reports, and support multimodal input, collaborative search planning, and streaming of intermediate steps.

Moonshot AI has released the Kimi K2.6 language model

Chinese startup Moonshot AI has released the Kimi K2.6 open language model, capable of running up to 300 subagents simultaneously.

The system distributes one major task between specialized assistants who can take thousands of steps to achieve a result, whether it's analyzing archived documents or creating a full-fledged website with animation.

The model is capable of working continuously for more than 12 hours and performing complex tool call chains in the programming languages Rust, Go and Python. When certain financial performance indicators for the use of the model are achieved, developers are asked to indicate its name in the interface of the final product.

Mind in wards: Scientists have developed a system for using AI in intensive care units across the country

A unique database will help predict the onset of critical condition in patients

OpenAI introduces the GPT-Rosalind model

OpenAI also introduced a specialized GPT-Rosalind model designed for pharmaceutical research. The tool has deep knowledge in chemistry, protein engineering and genomics, which should help scientists shorten the cycle of developing new drugs from the current 10-15 years.

Major biopharmaceutical companies such as Amgen and Moderna have already become partners of the project. The name of the model refers to the British biophysicist Rosalind Franklin, who was able to discover the double helix thanks to X-rays.

Anthropic has released Claude Opus 4.7

The company has released the Claude Opus 4.7 update, which deliberately limited the capabilities of AI in the field of cyber attacks. Despite the high performance in writing code and following instructions, the model was specially trained so as not to help in finding vulnerabilities in IT systems.

To do this, the developers have implemented a double protection: reducing the relevant skills at the training stage and automatic filters for blocking suspicious requests. More powerful security testing tools will be available only to verified specialists as part of a closed program.