Reflection AI Launches Asimov: Breakthrough Agent for Code Comprehension

Reflection AI has launched Asimov, its first product milestone and what it calls a research agent for code understanding. The company is pitching it less as a code writer and more as a tool for making sense of code that already exists. That’s the real slog for most engineers anyway: not typing lines, but tracing systems, decisions, and dependencies.

The pitch is blunt. Reflection says engineers spend only a minority of their time writing code and most of it understanding and designing code. So Asimov is built to act like a shared memory for an engineering team, pulling in entire codebases, architecture docs, GitHub threads, chat history, and more. The goal is to turn scattered knowledge into a single source of truth.

Reflection also says Asimov can capture the kind of tribal knowledge that usually lives in one senior engineer’s head. Teams can feed it updates such as “@asimov remember X works in Y way,” and the system uses permissioned role-based access control so organizations can decide who gets to edit that memory. In other words, it’s not just a personal assistant; it’s meant to be a team asset.

There’s some tech behind the pitch too. Asimov uses a multi-agent setup with many long-context retrievers and one short-context reasoning agent that combines the results into an answer. Reflection says the current version runs on third-party models, while the company is training its own models to improve performance over time.

The early claim is that this already works. In blind testing with maintainers of some large open-source projects, Asimov’s answers were preferred a majority of the time over Cursor Ask and Claude Code, including Sonnet 3.7 and 4. That’s the sort of comparison that gets attention fast, especially in a market where everyone is chasing the next useful AI coding tool.