How an AI Agent Works

Topics Covered

• the agent loop: an LLM API call, a tool the model asks for, the result sent back, repeated
• the context window: what the model reads on every request (instructions file, files read, the conversation)
• why a long chat gets slower and costlier, and a new chat per task
• the harness that steers an agent (the instructions file, PRD.md)
• when to let the agent run vs review each step
• safety (it can modify files and run commands)

Content Slides Open fullscreen ↗

Taught In

• Tuesday, 12/08 — Final demos, part 2; how an AI agent works; wrap-up