← back
Amp Code: Next Generation AI Coding – Beyang Liu, Amp Code
Takeaway
Effective coding agents come from curated tools, dedicated subagents for search/reasoning, and a review UI — not from piling on MCP servers.
Summary
- Amp is an opinionated frontier coding agent with a custom terminal UI and VS Code integration; Sourcegraph's bet is that MCP server proliferation hurts agents (context confusion) so Amp invests in a curated core toolset instead.
- Specialized subagents are key: Finder (fast codebase search with limited toolset), Oracle (reasoning model invoked on demand to keep main agent snappy), Librarian (external library docs).
- Subagents act like subroutines — preserving the main context window by returning only relevant results, addressing the 'doom loop' where agents over-grep and run out of context before editing.
- Bottleneck for the user is code review of agent output; Amp built a review UI to streamline reading agent diffs so devs can run more agents in parallel.
coding-agentssubagentstool-design
Original description
Introduction to Amp Code and its approach to AI-powered software development. Speaker: Beyang Liu | Co-founder & CTO, Amp Code https://x.com/beyang https://www.linkedin.com/in/beyang-liu/ https://github.com/beyang Timestamps: 00:00 Introduction & The "Weird" Ethos 01:19 Amp Terminal UI & Editor Integration 03:02 The "Review Bottleneck" & Review Interface 03:57 Defining an Agent: For Loops & Tool Calls 04:46 The Argument Against MCP (Context Confusion) 06:04 Tool Call Context Exhaustion 14:51 Business Model: Ads in the Terminal 15:38 Community & The "Weird" Builder Cohort