course.yaml

StartDate: 2025-12-02
EndDate: 2025-12-03

Slug: ai-agents-v2
Title: Build an AI Agent from Scratch
Subtitle: Two days building a general-purpose agent from first principles—tool calls, loops, memory, evals, and guardrails.

Teachers:
  - moss

Excerpt: >
  Build a general-purpose agent from scratch in TypeScript. No frameworks, no
  magic—just you, an LLM API, a tool-aware loop, and the patterns that make
  agents reliable across files, web search, code execution, and beyond.

Benefits:
  - Understand the core primitives of an agent: models, tools, history, memory, and orchestration
  - Implement a tool-calling loop with conversation-style messages instead of relying on a framework
  - Plug in filesystem tools, 3rd-party tools (web search, code execution, browser handoff), and other SDKs with minimal extra code
  - Manage context windows with summarization and retrieval so your agent can use local files and search results without blowing the token limit
  - Design and run evals that cover both single-step decisions and full multi-step runs
  - Add guardrails and human approvals around risky tools like shell commands or bulk edits
  - Treat agents as composable building blocks that can be called as tools or wired into other systems

Summary: >
  You’ll build a generic agent that can orchestrate tools on your machine:
  reading and transforming files, calling 3rd-party tools like web search and
  code execution, and handing results off to a local browser. The focus is the
  agent core: a loop that maintains conversation history, uses tool calling to
  select tools and arguments, updates messages based on tool results, and
  decides when to stop. Along the way, you’ll learn how to manage context with
  summarization and retrieval, layer in evals to catch failures, and add
  guardrails and human-in-the-loop checks for sensitive actions. By the end,
  you’ll have an agent you can keep extending with new tools, protocols, and
  interfaces.

Audience:
  Overview:
    - Frontend and UI engineers building apps in JavaScript or TypeScript
    - Full-stack or backend JS/TS devs who want to understand agents, not just wire up an SDK
    - Engineers who like tool-using editors and playgrounds and want a general agent they can customize
    - Anyone exploring agents as reusable components in their own products or internal tools
  Prerequisites:
    - Comfortable with TypeScript and Node.js
    - Can run a local dev server and manage environment variables
    - An API key for at least one modern LLM provider
    - Basic familiarity with the command line is helpful but not required

Schedule:
  # Day 1
  - 0930: "Intro to Agents"
  - 1030: "Break"
  - 1045: "Tool Calling"
  - 1200: "Lunch"
  - 1300: "Single Turn Evals"
  - 1415: "Break"
  - 1430: "The Agent Loop"
  - 1545: "End"

  # Day 2
  - 0930: "Multi-turn Evals"
  - 1030: "Break"
  - 1045: "File System Tools"
  - 1200: "Lunch"
  - 1300: "Web Search + Context Management"
  - 1400: "Break"
  - 1415: "Shell Tool"
  - 1515: "Break"
  - 1530: "HITL"
  - 1630: "End"