Co-founder and CTO of Amp Code, building next-generation AI coding agents with multi-agent architecture. Previously co-founded Sourcegraph, the enterprise code intelligence platform serving 4/5 FAANG companies.
Champion of Pragmatic AI Architecture
Beyang Liu advocates for practical AI systems that work reliably in production. His philosophy: the bottleneck isn’t model capability—it’s system architecture. Strong models need intelligent agent design, context management, and specialized reasoning paths to deliver production-grade results.
Current Work
As CTO of Amp Code, Beyang leads development of an AI coding platform built on multi-agent architecture featuring dual-model systems (Smart agents for deep reasoning, Rush agents for fast execution) and four specialized subagents that manage context efficiently. The platform leverages Sourcegraph’s code intelligence capabilities to power AI-driven code understanding.
Beyang actively shares technical insights on his blog at beyang.org, covering topics like configuring VS Code for agentic coding, developer productivity, and building developer tools. He also writes on Medium and runs “Tour de Source” on Substack.
Background
Previously co-founded Sourcegraph, building the enterprise code intelligence platform serving major tech companies including Uber, Stripe, and Coinbase. Stanford AI Lab researcher with published work in probabilistic graphical models and computer vision, holding B.S. and M.S. in Computer Science. Earlier software engineering experience at Palantir Technologies.
Philosophy on AI Coding
Specialization over generality - Different tasks benefit from different models and architectures; Amp’s dual-system approach reflects this reality.
Context is currency - Tool calls consume context; subagents compartmentalize concerns and prevent context confusion.
Architecture matters more than models - The limiting factor in production AI is system design, not raw model capability.
Editor as “readitor” - Modern development emphasizes understanding code before modification; AI shifts focus from writing to reviewing generated code.
Amp Code
Amp Code implements a four-subagent architecture that solves context management challenges in AI coding: Finder handles codebase search, Oracle provides careful reasoning and code review, Librarian manages library and dependency understanding, and Kraken performs refactoring and code transformation. Each subagent has focused context and specialized capabilities.
Conference Appearance
Event: AI Engineering Code Summit 2025 Date: November 21, 2025 Time: 2:25 PM - 2:44 PM Session: “Amp Code: Next-Generation AI Coding”
Presented multi-agent architecture for production AI coding, emphasizing context management, specialized reasoning agents, and the shift from code generation to code review workflows.