Reflection is taking a different route in the race to build more capable AI agents. Its new system, Asimov, is meant to learn how software is produced inside a company by reading not only code, but also the surrounding communication and documentation that explain why that code exists.
The bet is straightforward: if an AI agent can understand the full engineering context behind a product, it may become more useful than a tool that simply writes snippets on command.
A Software Agent Built Around Context
Asimov was developed by Reflection, a small startup cofounded by top AI researchers from Google. The company’s wider ambition is to build superintelligent AI, a goal that other major AI labs are also pursuing.
The agent is designed to ingest a broad set of internal engineering material. That includes code, emails, Slack messages, project updates, and other documentation. The purpose is to learn how those pieces connect to the finished software a team ships.
That makes Asimov different from many coding assistants. Its core job is not just to generate code. It is to understand the environment in which code is written, reviewed, changed, and explained.
Misha Laskin, Reflection’s CEO, argues that mastering software development is a natural path toward smarter AI agents. He previously worked on Gemini and agents at Google DeepMind, and he says coding gives AI a direct way to interact with the world. By contrast, he suggests that using human interfaces and browsing the web is not as natural for a large language model.
Why Reading Matters More Than Writing
Laskin says Asimov is intended to spend more time reading code than writing it. That emphasis matters because modern software work is rarely contained in the repository alone.
Engineering teams make decisions in messages, planning documents, updates, and discussion threads. A model that sees only source files may miss the reasons behind a design choice, the tradeoffs that were already debated, or the customer problem a feature is supposed to solve.
As Laskin put it, “Everyone is really focusing on code generation,” adding, “But how to make agents useful in a team setting is really not solved. We are in kind of this semiautonomous phase where agents are just starting to work.”
Asimov is built as a group of smaller agents working together. Some retrieve information. A larger reasoning agent then combines that information into a coherent answer to a user’s question.
That structure supports a practical use case: answering questions about a codebase and the engineering system around it. For teams with large projects, the value may come from faster understanding rather than faster typing.
The Performance Claim and the Skepticism
Reflection says Asimov is already preferred over some leading AI tools in certain tests. In a survey conducted by Reflection, developers working on large open source projects preferred Asimov’s answers 82 percent of the time, compared to 63 percent for Anthropic’s Claude Code running its model Sonnet 4.
That is a notable claim, but it is not the same as independent proof of broad superiority. Daniel Jackson, a computer scientist at Massachusetts Institute of Technology, says Reflection’s wider information gathering seems promising. He also says the benefits remain to be seen and that the company’s survey is not enough to convince him of broad benefits.
Jackson points to two important risks. One is cost, because reading and processing more material could increase computation demands. The other is security, because an agent with access to internal engineering context may also touch sensitive communications.
His concern is direct: “It would be reading all these private messages,” he says.
Reflection says Asimov runs inside customers' virtual private clouds, so customer data remains with the customer. The company also says it does not train on customer data.
Reinforcement Learning Behind the Approach
Reflection’s CTO, Ioannis Antonoglou, brings experience from Google DeepMind. He was a founding engineer there and worked on reinforcement learning, a technique famously used to build AlphaGo, the program that learned to play Go at a superhuman level.
Reinforcement learning trains an AI model through practice combined with positive and negative feedback. In large language models, it can help produce more coherent and pleasing answers. With more training, it can also support simulated reasoning, where difficult problems are broken into steps.
Asimov currently uses open source models. Reflection is also using reinforcement learning to post-train custom models that it says perform even better.
The training goal is not to win a board game. It is to learn how finished software gets built. Reflection uses data from human annotators and also creates synthetic data. According to the company, customer data is not used for training.
Antonoglou compares the idea to an AI research agent for engineering systems. “We've actually built something like Deep Research but for your engineering systems,” he says. He adds, “We've seen that in big engineering teams, a lot of the knowledge is actually stored outside of the codebase.”
From Engineering Assistant to Company Oracle
Reflection’s long-term vision is much larger than a better code assistant. Its leaders believe an increasingly intelligent agent could become an oracle for a company’s institutional and organizational knowledge.
From there, they see a path where the agent learns to build and repair software autonomously. Eventually, they believe such systems could invent new algorithms, hardware, and products autonomously.
That ambition sits inside a competitive AI market. Meta recently created a new Superintelligence Lab and has promised huge sums to researchers interested in joining the effort. Companies with deep resources are investing heavily in talent and infrastructure, which may make the race harder for startups like Reflection.
Reflection does have backing from Sequoia. Stephanie Zhan, a partner at the investment firm, says the startup “punches at the same level as the frontier labs.”
The near-term path may be more practical than the long-term vision. Laskin says customers have already asked whether technical sales staff or technical support teams could use the system. That suggests Asimov’s first wider role may be answering hard product and engineering questions for people who need technical context but are not necessarily writing code every day.