Virtual IT company with AI agents managed only a quarter of tasks

Post Thumbnail

Researchers from Carnegie Mellon University conducted an unusual experiment. Creating a virtual IT company where AI agents worked instead of people. The results turned out to be quite interesting, and now I’ll tell you what came out of it.

The team of scientists launched an environment called The Agent Company. A virtual software development company. In it, artificial intelligence agents had to perform all the same things that a regular IT specialist does. They had to write code, use the terminal, search for information in the browser, work with Google Drive. As well as communicate with colleagues in messenger and manage tasks in Jira and GitLab.

To objectively evaluate the work of artificial intelligence, the researchers applied a checkpoint system instead of a simple “completed/not completed.” The cost of each model’s work was also taken into account by the number of tokens used. The most modern language models participated in the experiment. Claude 3.5 Sonnet, Gemini-2.0 Flash, GPT-4-o, Gemini-1.5-Pro, Llama-3.3 and 3.1, Qwen-2.5 and others.

In total, the agents were given 175 tasks from different areas. Software development, project management, HR and other directions. According to estimates, it would take about 3000 hours for 20 live people to complete this amount of work.

And what happened? Even the best of the participants, Claude 3.5 Sonnet, managed only 1/4 of the basic responsibilities! Gemini 2.0 Flash came in second place, but already with a large gap. Only 11.4% of completed tasks. All other models showed a result of less than 9%. How do you like the results?

Interestingly, working with GitLab and programming came quite easily to artificial intelligence. But the most difficult tasks turned out to be quite simple things. Filling out forms, navigating websites, scheduling meetings, and communicating in messenger. Surprisingly.

It turns out that despite all the achievements of artificial intelligence, it is still very far from fully replacing people in office work. Paradoxically, tasks that people consider difficult, like programming, for artificial intelligence turned out to be easier than everyday affairs. Which a person handles on autopilot.

AIvengo > Reviews > Virtual IT company with AI agents managed only a quarter of tasks
Почитать из последнего
UBTech will send Walker S2 robots to serve on China's border for $37 million
Chinese company UBTech won a contract for $37 million. And will send humanoid robots Walker S2 to serve on China's border with Vietnam. South China Morning Post reports that the robots will interact with tourists and staff, perform logistics operations, inspect cargo and patrol the area. And characteristically — they can independently change their battery.
Anthropic accidentally revealed an internal document about Claude's "soul"
Anthropic accidentally revealed the "soul" of artificial intelligence to a user. And this is not a metaphor. This is a quite specific internal document.
Jensen Huang ordered Nvidia employees to use AI everywhere
Jensen Huang announced total mobilization under the banner of artificial intelligence inside Nvidia. And this is no longer a recommendation. This is a requirement.
AI chatbots generate content that exacerbates eating disorders
A joint study by Stanford University and the Center for Democracy and Technology showed a disturbing picture. Chatbots with artificial intelligence pose a serious risk to people with eating disorders. Scientists warn that neural networks hand out harmful advice about diets. They suggest ways to hide the disorder and generate "inspiring weight loss content" that worsens the problem.
OpenAGI released the Lux model that overtakes Google and OpenAI
Startup OpenAGI released the Lux model for computer control and claims this is a breakthrough. According to benchmarks, the model overtakes analogues from Google, OpenAI and Anthropic by a whole generation. Moreover, it works faster. About 1 second per step instead of 3 seconds for competitors. And 10 times cheaper in cost per processing 1 token.