Reflection AI Launches Asimov: Breakthrough Agent for Code Comprehension

Asimov targets code understanding, not writing — where engineers actually spend time.
Blind tests: Asimov preferred over Cursor Ask and Claude Code (Sonnet 3.7 + 4).
- Open source project maintainers were the evaluators.
Ingests full codebases, docs, GitHub threads, chat history → persistent org memory.
Tribal knowledge capture: @asimov remember X works in Y way offloads senior engineers.
- Role-based access control included.
Multi-agent: many small long-context retrievers + one large short-context combiner.
- Currently third-party models; proprietary training in progress.
Sequoia thesis: best products couple model intelligence + application layer.
- Cited Deep Research and Claude Code as proof points.

Stephanie Zhan and Charlie Curnin (Sequoia Capital) · 2025-07-16 · Read on sequoiacap.com