Beyang Liu co-founded Amp, the independent coding-agent company, and Sourcegraph, the code-search company he previously led as chief technology officer. His work addresses a foundational challenge for AI-assisted programming: giving developers and their agents enough understanding of complex software to change it intelligently.
Liu studied computer science at Stanford, including machine-learning research in probabilistic graphical models and computer vision, before working at Palantir on data-analysis software for large companies and banks. In 2013, he and Quinn Slack founded Sourcegraph, initially focused on making sprawling codebases searchable. The company expanded into code intelligence, large-scale code changes, and Cody, an AI coding assistant that grounds its responses in relevant material retrieved from across an organization’s software.
His writing on developer productivity distinguishes the broader software-delivery process from the developer’s immediate cycle of reading code, making changes, testing, and iterating. That systems-level perspective rejects simplistic proxies such as lines written: productivity depends on whether engineers can understand their environment, maintain momentum, and deliver useful work.
As models progressed from autocomplete and retrieval-assisted chat to autonomous tool use, Liu helped build Amp around agent-native workflows. In December 2025, he and Slack launched Amp as an independent company, while Dan Adler became Sourcegraph’s chief executive. Liu remains on Sourcegraph’s board.
- Codebase context determines agent usefulness. Effective assistance requires finding relevant implementations, dependencies, conventions, and architectural constraints beyond the currently open file. Liu treats retrieval as an engineering problem with two competing failure modes: exhausting the agent’s context window through excessive searching, or gathering so little information that it repeatedly attempts the wrong fix.
- Specialized subagents preserve working context. Amp delegates bounded tasks to separate agents: Finder searches codebases, Oracle handles difficult reasoning, Librarian examines external dependencies, and the experimental Kraken targets large refactors. Each returns concentrated results to the primary agent. Liu favors tightly integrated tools and matched models over generic integrations or a user-facing model picker, as he explains in his account of Amp’s architecture.
- Agent evaluation should measure organizational outcomes. Liu assesses coding tools through controlled comparisons, carefully interpreted usage data, developer feedback, and concrete engineering priorities such as delivery speed, test coverage, code review, or modernization. Generated lines and accepted completions cannot, by themselves, establish software quality or business value.
- Human accountability depends on fast feedback loops. Detailed instructions, tests, browser automation, diagnostics, and diff-based review help agents correct their own work while keeping developers responsible for shipped code. In a production-code demonstration, an apparently simple interface change required exposing the appropriate configuration field without leaking administrative secrets—a practical example of why plausible generated code is insufficient without architectural understanding.
