A new Training for AI agents can learn how to build software by filling in company data and knowing how to lead to the final product, which can be both a more capable software assistant and a small step towards smarter AI.
The new agency, called Asimov, was developed by Reflection, a small but ambitious startup that is confused by top Google researchers. Asimov reads code as well as emails, slack messages, project updates and other documents with the goal of learning all of this how to generate completed software together.
The ultimate goal of reflection is to build super intelligent AI, which other leading AI labs say they are working on. Meta recently created a new super smart lab that provides huge funding to researchers interested in joining its new work.
I visited the Brooklyn neighborhood of Reflection, headquartered in Williamsburg, NY, right across from the top pickle club to learn how to get to the super smart reflection program before the game.
The company’s CEO Misha Laskin said the ideal way to build super AI agents is to get them real chief code because it’s the easiest and most natural way for them to interact with the world. While other companies are building agents that use human user interfaces and browse the web, Laskin, who previously worked on Gemini and agents at Google DeepMind, said that this is almost unnaturally a large language model. Teaching AI to understand software development will also produce more useful coding assistants, Laskin added.
Asimov aims to spend more time reading code than writing code, Ruskin said. “Everyone is really focused on code generation,” he told me. “However, how to make an agent useful in a team environment is not solved. We are in this semi-autonomous phase and agents are just starting to work.”
Asimov actually consists of several smaller agents inside the windbreaker. Agents all work together to understand the code and answer users’ questions about it. The smaller proxy retrieves information, and a larger inference agent combines this information into a coherent answer to the query.
Reflection claims that with certain measures, Asimov has been considered to outperform some leading AI tools. In a survey conducted by reflection, the company found that developers working on large open source projects asked questions 82% of the time, their answers took precedence over 82% of the time, while anthropomorphic Claude code was 63%, running its model sonnet 4.
Daniel Jackson, a computer scientist at Massachusetts Institute of Technology, said that given the wider scope of its information collection, the reflection approach seems promising. However, Jackson added that the benefits of the approach remain to be seen and the company’s investigation is not enough to convince him of the wide range of benefits. He noted that the method could also increase computational costs and could cause new security issues. “It’s going to be reading all this private message,” he said.
Reflection says that multiple approaches can reduce computing costs and leverage a secure environment that provides security compared to some conventional SaaS tools.
