Reflection AI, a startup founded by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, says that understanding software, and not just generating it, is essential to building more intelligent AI systems. They’ve designed their first product, Asimov, to do just that. The company describes it as a code comprehension agent: a tool built to help engineering teams navigate complex codebases and surface relevant knowledge buried across systems.
Asimov is also Reflection’s first step toward a larger goal: building superintelligent agents capable of reasoning across domains. “Superintelligent code understanding is the prerequisite for superintelligent code generation,” the company wrote in a blog post announcing the product. The underlying belief is that software is one of the most structured, information-dense environments in which to train and evaluate advanced AI.
For now, Asimov is focused on more immediate use cases: helping engineers understand how large systems work, preserving institutional memory, and making technical knowledge accessible across teams.
A Teamwide System of Record
The company says Asimov is built to integrate directly into an organization’s development environment. It ingests codebases, documentation, project threads, Slack logs, and more, creating what Reflection calls a persistent memory of the software system. The architecture uses multiple AI agents: smaller retrievers gather relevant context, while a central reasoning model synthesizes responses to user queries.
The agent can recall prior decisions, trace dependencies across files, and return answers with citations to the original sources, such as commits, design docs, or internal conversations. Engineers can annotate Asimov’s knowledge with prompts like “@asimov remember X works in Y way,” which allows senior developers to document informal decisions in a structured, shareable format. A role-based access system governs who can add or update knowledge.
According to Reflection Asimov has been evaluated in blind testing by maintainers of large open-source projects, who preferred its responses over alternatives like Claude Code and Cursor Ask in the majority of comparisons. These early results have not been benchmarked against models from OpenAI or GitHub, and the company has not published detailed evaluation data.
Designed with enterprise deployment in mind, Asimov runs inside a customer’s virtual private cloud and integrates with major cloud providers. The company says its design choices reflect concerns about data control and internal security, particularly when ingesting large volumes of proprietary communication and source code.
Still Building for Near-Term Utility
Both founders previously worked on reinforcement learning systems at DeepMind. Antonoglou was a key contributor to AlphaGo, and they bring that background into their approach to software agents. Asimov uses reinforcement learning techniques alongside supervised training and synthetic data generation. The current release runs on third-party models, but the company is developing its own.
While many AI tools in the software space are built around code generation: autocomplete-style features that assist while writing, Asimov is meant to function more like a researcher or archivist. “Everyone is really focusing on code generation,” Laskin told Wired. “But how to make agents useful in a team setting is really not solved.”
Reflection sees that gap between individual productivity tools and teamwide systems as an opportunity. The company believes that building agents with deep, organization-specific memory and reasoning ability could lay the groundwork for more autonomous systems in the future. In theory, a system that understands the full context of how software works could one day build or repair it with limited oversight.
In March, Reflection raised $130 million from investors including Sequoia, Lightspeed, CRV, and NVIDIA. The company’s valuation now stands at $555 million. It says the funding will go toward model development and hiring.
Asimov is not yet broadly released. In the short term, Reflection is positioning it as a way to improve engineering team efficiency and preserve knowledge as teams scale. Whether that becomes a stepping stone to more general AI capabilities will depend on how well Asimov performs in real-world environments and whether businesses find its approach useful enough to adopt.
