Reflection AI's Asimov: A New Breed of AI Agent Revolutionizing Software Development

Former Top Google Researchers Have Made A New Kind of AI Agent

A new kind of artificial intelligence agent, trained to understand how software is built by gorging on a company's data and learning how this leads to an end product, could be both a more capable software assistant and a small step towards much smarter AI. The new agent, called Asimov, was developed by Reflection, a small but ambitious startup confounded by top AI researchers from Google. Asimov reads code as well as emails, Slack messages, project updates and other documentation with the goal of learning how all this leads together to produce a finished piece of software. Reflection's ultimate goal is building superintelligent AI -- something that other leading AI labs say they are working towards. Meta recently created a new Superintelligence Lab, promising huge sums to researchers interested in joining its new effort. I visited Reflection's headquarters in the Brooklyn neighborhood of Williamsburg, New York, just across the road from a swanky-looking pickleball club, to see how Reflection plans to reach superintelligence ahead of the competition. The company's CEO, Misha Laskin, says the ideal way to build supersmart AI agents is to have them truly master coding, since this is the simplest, 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, says this hardly comes naturally to a large language model. Laskin adds that teaching AI to make sense of software development will also produce much more useful coding assistants. Laskin says Asimov is designed to spend more time reading code rather than writing it. "Everyone is really focusing on code generation," he told me. "But how to make agents useful in a team setting is really not solved. We are in kind of this semi-autonomous phase where agents are just starting to work." Asimov actually consists of several smaller agents inside a trench coat. The agents all work together to understand code and answer users' queries about it. The smaller agents retrieve information, and one larger reasoning agent synthesizes this information into a coherent answer to a query. Reflection claims that Asimov already is perceived to outperform some leading AI tools by some measures. In a survey conducted by Reflection, the company found that developers working on large open source projects who asked questions preferred answers from Asimov 82 percent of the time compared to 63 percent for Anthropic's Claude Code running its model Sonnet 4. Daniel Jackson, a computer scientist at Massachusetts Institute of Technology, says Reflection's approach seems promising given the broader scope of its information gathering. Jackson adds, however, that the benefits of the approach remain to be seen, and the company's survey is not enough to convince him of broad benefits. He notes that the approach could also increase computation costs and potentially create new security issues. "It would be reading all these private messages," he says. Reflection says the multiagent approach mitigates computation costs and that it makes use of a secure environment that provides more security than some conventional SaaS tools.

Reflection AI's autonomous coding agent Asimov learns from more than just code - SiliconANGLE

Reflection AI's autonomous coding agent Asimov learns from more than just code Artificial intelligence startup Reflection AI Inc. is looking to change the way AI agents are designed and built, with a view to achieving so-called "superintelligence" much faster. The company has developed an autonomous agent known as Asimov, which has been trained to understand how software is created by ingesting not only code, but the entirety of a business's data to try and piece together why an application or system does what it does. In an interview with Wired.com, Reflection AI co-founder and Chief Executive Misha Laskin said that Asimov reads everything from emails to slack messages, project notes to documentation, in addition to the code, to learn everything about how and why the app was created. He explained that he believes this is the simplest and most natural way for AI agents to become masters at coding. It's a stark contrast to how AI agents work at most other AI companies, where they tend to be more focused on code generation than actually understanding the software they help to build. According to Laskin, the goal is to make AI agents more useful in a team setting. "We are in this kind of semi-autonomous phase, where agents are just starting to work," he said. Asimov is actually a collection of multiple smaller AI agents that are deployed inside customer's cloud environments so that the data remains within their control. Asimov's agents then cooperate with one another to try and understand the underlying code of whatever piece of software they've been assigned to, so they can answer any questions that human users might have about it. There are several smaller agents designed to retrieve the necessary data, and they work with a larger "reasoning" agent that collects all of their findings and tries to generate coherent answers to user's questions. Reflection AI's other co-founder, Chief Technology Officer Ioannis Antonoglou previously served as a founding engineer at Google LLC's AI research unit DeepMind, where he worked with Laskin on a project that sought to teach AI models to reason and play games. He was part of the team that developed AlphaGo, an AI system that learned how to play the ancient board game Go and defeat some of the world's best players. To teach AlphaGo, Antonoglou developed a groundbreaking training technique known as reinforcement learning, where AI models are taught their desired behavior using positive feedback to reward correct responses, and negative feedback when they make a mistake. The technique has since been implemented in large language model training, helping models to generate more coherent responses to user prompts. Reflection AI has applied reinforcement learning to Asimov, but instead of teaching it how to play Go, it's focused on enhancing its proficiency in coding. And by feeding it with not only code, but also emails, messages and other documents from those who built the software, the company provides Asimov with much more context about it. By combining these methods, Reflection AI believes the agent will eventually be able to create the same software independently. Antonoglou said the approach is not too dissimilar from Google's Deep Research agent, which scours hundreds or even thousands of sources from across the web to create comprehensive reports on almost any subject. "We've built something like Deep Research, but it's for your engineering systems," he said. "We've seen that in big engineering teams, a lot of the knowledge is actually stored outside of the codebase." The results so far have been encouraging, at least if the company itself is to be believed. An internally-conducted survey found that developers of large open-source software projects showed a preference for Asimov's responses and code outputs to 82% of their prompts. In contrast, they only liked the responses from Anthropic PBC's Claude Sonnet 4 model 63% of the time. Of course, that research is far from comprehensive, as there is no comparison with any of OpenAI's models, or those from coding-specific models such as GitHub's CoPilot or Cognition AI Inc.'s Devin, for example. Daniel Jackson, who is a professor of Computer Science at the Massachusetts Institute of Technology, told Wired that Asimov is a promising idea, and said he likes the idea of gathering a much broader set of data to inform its coding outputs. But he said it's still unproven, and he also raised concerns about privacy, noting that it will be "reading all these private messages" between developers and their colleagues. In any case, Reflection AI believes it's on the right track, and says it ultimately sees Asimov evolving to become a kind of coding "oracle" for businesses, leveraging its comprehensive knowledge of everything they do to autonomously build new software and products on their behalf.

Meet Asimov: Reflection's AI agent to help write the best software code for anyone

Software development thrives on creativity, collaboration, and deep understanding, yet it's often bogged down by the complexity of sprawling codebases and the unwritten "tribal knowledge" held by a select few. Reflection AI's Asimov is here to change that narrative. This code research agent goes beyond generating lines of code, it dives into the heart of software ecosystems. It absorbs everything from code repositories to architecture documents, GitHub threads, emails, and Slack conversations. By building a comprehensive knowledge base that captures not just the "what" but the "why" behind technical decisions, Asimov is revolutionizing how engineering teams operate, making software development faster, smarter, and more accessible to everyone from developers to support staff. Also read: OpenAI, Google, Anthropic researchers warn about AI 'thoughts': Urgent need explained Traditional AI coding tools often focus on churning out code, sometimes missing the nuanced context that defines robust software. Asimov, however, is designed to understand. It ingests entire codebases and supplementary materials, creating a holistic view of a project's technical and business logic. Its multi-agent architecture is key: smaller "retriever" agents scour diverse data sources, code, documentation, communication threads, while a central "reasoning" agent synthesizes this information into clear, contextually accurate responses. In blind tests against Claude and Cursor, Asimov's answers were preferred 60-80% of the time by users. For developers wrestling with legacy systems or intricate dependencies, Asimov slashes the time spent by up to 70%, deciphering code, allowing them to focus on building innovative solutions. Its ability to recall the reasoning behind past decisions also ensures that critical insights remain accessible, even as teams evolve. Asimov isn't just powerful, it's practical. Designed with enterprise security in mind, it can be deployed within a company's virtual private cloud (VPC) through partnerships with AWS, Microsoft, and Google. A permissioned Role-Based Access Control (RBAC) system allows teams to manage who can update Asimov's knowledge base. They are maintaining strict control over proprietary information. Also read: Mark Zuckerberg says AI will write most of Meta's AI code by 2026 Engineers can easily enrich its memory with commands like "@asimov remember X works in Y way," effectively distributing senior developers' expertise across the organization. This scalability extends beyond engineering teams. Technical sales and support staff can tap into Asimov's insights without burdening developers, streamlining workflows and fostering cross-departmental collaboration. By reducing the cognitive load on engineers and preserving institutional knowledge, Asimov empowers organizations to operate more efficiently, regardless of size or complexity. Reflection AI, founded by ex-Google researchers, sees Asimov as a stepping stone to superintelligent AI. Trained with reinforcement learning and human feedback, it leverages third-party models but is being enhanced with custom models for more performance. The long-term vision is transformative: an AI "oracle" capable of not only answering complex queries but autonomously building, repairing, and innovating software, potentially inventing new algorithms or products. Asimov's ability to retain and share critical knowledge ensures that no decision is lost to turnover or time. This makes it an indispensable asset for modern software teams. Reflection AI, founded by ex-Google researchers, positions Asimov as a step toward advanced AI systems. Using reinforcement learning and human feedback, it currently relies on third-party models while developing custom ones for improved performance. The long-term goal is an AI capable of answering complex queries and handling software tasks, such as building code. Those interested in exploring Asimov can join the waitlist on Reflection AI's website. As a tool under development, its full capabilities are yet to be publicly tested. However, it shows promise to support software development by enhancing efficiency and accessibility for diverse teams.