Descriptions:
Beyang Liu, co-founder of Amp Code, presents the design philosophy and architecture behind Amp — a terminal-first, opinionated AI coding agent built for developers who want to run multiple agents in parallel. The talk is framed as a series of deliberate, contrarian architectural decisions made while building the product, starting with the foundational view that an agent is simply a for-loop with tool calls and a model — and that the real design surface is which model, which tools, and how they iterate together.
Amp’s most distinctive choice is replacing the model selector (a standard UX component in tools like Cursor and Windsurf) with a fixed set of specialized sub-agents: the Finder, a small fast model for codebase context discovery; the Oracle, a reasoning model invoked only for hard debugging or nuanced planning; the Librarian, which fetches context from external libraries and frameworks; and the Kraken, an experimental agent that writes codemods for large-scale refactors rather than editing files one at a time. Liu argues that offering model choice creates cognitive overhead and prevents deep optimization for any single model’s strengths.
The terminal UI was built from scratch on a custom framework to surface just enough information — streamed diffs, CLI commands, editor diagnostics from Emacs, Neovim, and JetBrains — without overwhelming users. Amp also ships a code review interface Liu identifies as the real bottleneck to parallelizing agent work: most developer time is now spent reviewing agent output, not writing code, and streamlining that review loop is what enables running two or three agents simultaneously.
📺 Source: AI Engineer · Published December 22, 2025
🏷️ Format: Keynote Launch
