Harnessing the Model: What Coding Agents Are Really Made Of
summary
Walks through the Open Code codebase and proves a coding agent is mostly harness not magic: TypeScript async/generator loops glue an LLM to tools, streaming token yields let the agent start tool calls before a full response, and the demo executes read package.json via Vercel AI SDK calling the OpenAI API and surfaces results in a debugger. Shows breakpoints, OpenTelemetry/Jaeger traces, and why context packing, tool descriptions, and verification loops (tests/linters) control hallucinations, session drift, and reliable fixes.
AI-written summary · about
chapters (12)
- 0:00Introduction and speaker background
- 1:11Why this talk about coding agents
- 2:13Exploring Open Code's architecture
- 7:53Why TypeScript for coding agents
- 8:36Core components of coding agents
- 11:49Generators in 30 seconds
- 19:08Generators' role in coding agents
- 20:51Open Code's use of generators
- 25:12Live demo setup
- 26:18Debugging Open Code's execution flow
- 35:18OpenTelemetry tracing insights
- 36:43Q&A and closing remarks
