Skip to main content

19 lessons · a working notebook

DeepSeek Harness, one task at a time

For developers already working with Codex, Claude Code, or similar agents. Study the execution loop, tool and policy boundaries, context pressure, recovery, and orchestration through reviewable repository work.

Start with the first lesson →Explore the syllabus ↓

180 minutes of reading. Leave time to run the exercises; this is not a timed course.

A notebook task moves through a local harness, a custom workspace, and a checked result.

What you’ll be able to do

  • Compare agent runtimes by execution semantics, extension seams, and evidence, not by prompt demos.
  • Design explicit boundaries for tools, permissions, context, sessions, and model routing.
  • Build a project-aware workbench with typed tools, scoped instructions, and a controlled workflow.
  • Recover uncertain runs and finish a capstone with reproducible checks and a reviewable handoff.

Before you start

  • Active experience with Codex, Claude Code, or another coding agent, including reviewing its diffs and tool calls.
  • A disposable local project, a supported DeepSeek Harness checkout, and enough JavaScript or TypeScript to inspect examples.
  • An authorized provider/model route for optional live exercises. Local checks can run without model access.

The syllabus

Read in order, or return to the task you need.