Ex-Google AI Researchers Develop Advanced Coding Agent Asimov

A new artificial intelligence agent named Asimov, developed by startup Reflection, aims to revolutionize how AI understands software development. Founded by former top Google AI researchers, Reflection seeks to push the boundaries of AI by training models to comprehend not just code but also the broader context of software creation, including emails, Slack messages, and project documentation.

The Vision Behind Asimov

Reflection's CEO, Misha Laskin, argues that mastering coding is the most natural path to superintelligent AI. Unlike other AI tools that interact with human interfaces or browse the web, Asimov is designed to deeply understand software development. "Everyone is really focusing on code generation," Laskin says. "But how to make agents useful in a team setting is really not solved."

How Asimov Works

Asimov is a multi-agent system where smaller agents retrieve information and a larger reasoning agent synthesizes responses. Reflection claims that in a survey, developers preferred Asimov's answers 82% of the time over Anthropic’s Claude Code (63%). However, Daniel Jackson, a computer scientist at MIT, cautions that the broader benefits and potential security risks of such an approach remain unproven.

Reinforcement Learning and Future Goals

Reflection’s CTO, Ioannis Antonoglou, a founding engineer at Google DeepMind, applies reinforcement learning—a technique used in AlphaGo—to train Asimov. The agent learns by breaking problems into steps and receiving feedback, similar to how humans learn. Reflection is also post-training custom models to improve performance.

Competitive Landscape

With giants like Meta investing heavily in superintelligence, startups like Reflection face stiff competition. However, Stephanie Zhan of Sequoia, a Reflection backer, believes the startup "punches at the same level as the frontier labs."

The Road Ahead

Reflection envisions Asimov evolving into an autonomous oracle for institutional knowledge, capable of inventing new algorithms and products. For now, the focus is on practical applications, such as assisting technical sales and support teams.

"We've actually been talking to customers who’ve started asking, can our technical sales staff, or our technical support team use this?" Laskin says.