Amp Code: Next-Generation AI Coding

2:25pm - 2:44pm | Amp Code: Next-Generation AI Coding

Speaker: Beyang Liu, Co-founder & CTO, Amp Code / Sourcegraph

Speaker Profile: Full Speaker Profile

Bio: Co-founder & CTO, Amp Code / Sourcegraph

Topic: Introduction to Amp Code and its approach to AI-powered software development

Notes

• quiry
• connects to emacs or whatever other editor
• wrote their own tui
• “having found the motivations yet to fork vs code”
• most of the time doing code review
  • (smarter linting)
• why is this different

  • mcp or customer tools

    • focus most of our attention on the internal tools calls of amp
    • avoiding context confusion with having too many mcps tools
    • tool calls themselves eat up context (context management again!)
    • subagents are the solve to clean up the context messiness
      • conserve and extend the context
      • 4 major subagents
        • finder -> codebase search
        • oracle -> reasoning, careful review (this is how amp does reasoning)
        • librarian -> library use
        • kraken -> refactoring model

  • models vs agents

    • smart agent
      • oracle
      • librarian
      • finder
    • rush
      • quick path
    • between intelligence and speed

  • to reason to not reason

    • only switch the main model a few days ago

  • gui vs tui

    • we are doing both
    • an editor is more of a “readitor”
    • very fleshed out diff editor and how to explore what you see

  • smart or face

  • is coding still a craft

  • smart or cheap

    • rush models aren’t free but close
    • costs
    • the have ads inside!

• We need to relearn the craft of how to code together
  • the ability to share threads with each other
  • there’s a link here to midjourney/discord that is really interesting
• buildcrew.team

Slides

Slide: 14-25

Key Point: Amp uses a hybrid approach, combining built-in custom tools for core coding functionality (file operations, search, code understanding) with MCP (Model Context Protocol) for specific integrations like browser automation via Playwright.

Literal Content:

• Title: “To MCP or Not To MCP?”
• Two columns on black background:
  • Left: Built-in (checkmarks) - Lists tools like Bash, create_file, edit_file, finder, format_file, get_diagnostics, glob, Grep, librarian, mermaid, oracle, read, read_mcp_resource, read_thread, read_web_page, Task, todo_read
  • Right: Playwright (MCP icon) - Lists MCP Playwright tools like mcp_playwright_browser_click, close, console_messages, drag, evaluate, file_upload, fill_form, handle_dialog, hover, install, navigate, navigate_back, network_requests, press_key, resize, run_code, select_option

Slide: 14-27

Key Point: Amp uses specialized sub-agents optimized for different tasks, each with tailored model selection (balancing speed vs. reasoning capability) and specific tool access. This demonstrates a multi-agent architecture approach.

Literal Content:

• Four panels with dramatic sci-fi imagery, each labeled with an agent name:
  1. Finder (Job: Codebase Search; Models: Sonnet 4.5 → Qwen3 → Haiku 4.5; Tools: Read, Grep, Glob)
  2. Oracle (Job: Reasoning, Tricky Bugs, Careful Review; Models: o3 → GPT-5.1; Tools: Read, Grep, Glob, Web Search, Read Amp Threads)
  3. Librarian (Job: Library Use; Models: Sonnet 4.5; Tools: Remote repository search API)
  4. Kraken (Job: Refactoring; Models: Sonnet 4.5; Tools: fastmod, etc.)
• AIE CODE logo in bottom right

Slide: 14-29

Key Point: Amp uses a routing architecture that directs tasks to different specialized agents. There’s a key decision between “smart” mode (using sophisticated agents like Oracle, Librarian, Finder) and “rush” mode (for faster execution), each decomposing into sub-tasks.

Literal Content:

• Title: “Amp’s Architecture”
• Diagram showing routing between different agents:
  • oracle (blue, left) - branches to three smaller nodes
  • librarian (magenta, center-left) - branches to three smaller nodes
  • finder (orange, center) - branches to two smaller nodes with “edit, grep, bash” label
  • smart (green, top-right) - single large node branching to three smaller nodes
  • rush (red, right) - single node branching to three smaller nodes
• Labels indicate “smart” and “rush” as routing options

Slide: 14-35

Key Point: Real users (including notable developers like Mitchell Hashimoto) are praising Amp’s unique features, particularly the ability to share session threads and its effectiveness as a coding agent. This slide provides social proof and community validation.

Literal Content:

• Title: “Join Us at the Frontier”
• Pink background
• Two Twitter/X posts shown:
  1. Mitchell Hashimoto (@mitchellh): “My favorite part about @AmpCode is that you can share your whole session globally. PRs with Amp threads attached make me very, very happy as a maintainer, here is one from a bug fix this morning:” - Shows embedded Amp thread about “Optimize SplitTree encoding with custom Codable” from ghostty.org/ghostty
  2. Kamil Husein (@KamilHusein): “My current fav coding agent is @AmpCode I finally invested in reading the manual and took some notes” - Shows “Amp RTFM Notes” from hamel.dev

Slide: 14-36

Key Point: Amp is building a developer community called “Build Crew” with incentives ($100 credit), gamification (XP, badges), and challenges to encourage adoption and engagement with their AI coding agent platform.

Literal Content:

• Title: “Join Us at the Frontier”
• Main content box titled “Build Crew”:
  • Description: “A community of devs shipping with agents. Get $100 in Amp credit, join a Discord community, take on build challenges, earn XP, and unlock badges.”
  • Button: “Join Build Crew →”
  • Footer text: “build challenges, XP system, private community”
• QR code on right side with URL: buildcrew.team
• Background image shows artistic/fantasy figure