This guide takes you from the fundamentals of Harness engineering to its daily practice with the two tools that dominate the market in May 2026: OpenAI Codex and Anthropic Claude Code. If you have arrived at this guide after experimenting with an agent and discovering that the results swing between the dazzling and the exasperating, you are in the right place. That oscillation is not a defect of the model: it is the unmistakable symptom that a harness is missing around it.
The guide covers exclusively the external harness: the one that you, as a Codex or Claude Code user, build on top of the infrastructure that Anthropic and OpenAI already provide. The internal harness (the query() loop, atomic file editing, handling the Hulk Problem in OpenAI's Responses API, dynamic prompt caching, etc.) is the vendors' responsibility; it exists and it is fascinating, but it falls outside the scope of this guide.
Table of contents
- Part I · Fundamentals
- Part II · Setting up the environment
- Part III · Hands-on practice with OpenAI Codex
- Part IV · Hands-on practice with Anthropic Claude Code
- Part V · Advanced patterns
- Part VI · Evaluation and continuous improvement
- Part VII · From hobbyist to Harness Engineer
- Glossary
- Annexes
-
Annex A: AGENTS.md template (multi-stack)
-
Annex B: CLAUDE.md template (multi-stack)
-
Annex C: Codex ↔ Claude Code comparison table
-
Annex D: harness audit checklist
-
If you want to learn at your own pace, with your Agile Club membership you can train and measure your knowledge of this topic in Skill Arena.
