From Code Search to AI Agents: Inside Sourcegraph's Transformation with CTO Beyang Liu

What It Covers

Sourcegraph CTO Beyang Liu discusses the company's evolution from code search to AI coding agents, revealing how Chinese open-source models dominate agentic tool use while American policy creates regulatory barriers. Liu explains why developers now orchestrate agents rather than write code, and how US AI safety narratives may be undermining national competitiveness.

Key Questions Answered

• •Agent Architecture Over Models: Sourcegraph treats models as implementation details within agent systems, not the core product. The same model with different system prompts and tool descriptions produces completely different behaviors. They run specialized sub-agents for tasks like context retrieval and debugging, each optimized on the Pareto frontier of intelligence versus latency, with some using single-digit billion parameter models for targeted edits.
• •Dual-Tier Agent Strategy: Sourcegraph deploys two top-level agents—a smart agent using frontier models like Claude Sonnet or GPT-5 for complex tasks, and a fast agent using smaller models that runs on advertisement-supported free tier. This approach recognizes that different workflows require different trade-offs: comprehensive features need maximum intelligence, while quick targeted edits optimize for speed once quality thresholds are met.
• •Chinese Model Dominance in Open Source: All competitive open-weight models for agentic tool use currently originate from China, including QuantFreeCoder and GLM. American companies produce open models, but their tool-calling capabilities fall short for production agent applications. Sourcegraph hosts all models on American servers for security, but post-trains exclusively on Chinese base models because US alternatives lack robust agentic capabilities at equivalent capability levels.
• •Developer Role Transformation: Developers now spend over ninety percent of time orchestrating agents and reviewing generated code rather than writing line-by-line implementations. The bottleneck shifted from code production to human comprehension—understanding agent outputs and making architectural trade-offs that only humans can evaluate. Traditional code review interfaces designed for human-paced development create friction when reviewing agent-generated changes spanning hundreds of files.
• •Policy Creates Competitive Disadvantage: US regulatory uncertainty around AI safety, copyright litigation, and state-by-state patchwork regulations make American companies reluctant to release competitive open-weight models. The Terminator-style AGI narrative, now dismissed by practitioners who use LLMs daily, persists in policymaking circles and drives risk-averse regulations. This regulatory overhang allows China to dominate open-source AI while America invented the entire ecosystem.