A conference talk transcript demystifying how coding agents like Claude Code work under the hood. Covers the core agent loop (while loop + tool descriptions + conversation history), how models access local files via tool calls, MCP endpoints as agent-friendly APIs, Claude Code's skills system, and hooks for extending agent behavior. The speaker also built a custom open-source 'Coding Agent Explorer' tool to intercept and visualize HTTP traffic between the agent and the model, and demonstrates building a minimal coding agent in under 200 lines of C#. Key insight: every request includes a system prompt (~25K chars), tool descriptions (~90K chars), and full conversation history — making context management critical for cost and reliability.
•44m watch time
212 Impressions