02 · Part 2

 The same lens on Claude Code .

 We just extracted an architecture from Codex. Real test of a good abstraction: does it explain a competing product we didn’t derive it from? Use the shared vocabulary — not a feature-by-feature marketing comparison.

 Side by side

 One architecture, two products

 Architecture 
 What differs? 
 Discussion 

 Sort it out

 Codex, Claude Code — or both?

 For each statement, decide whether it describes Codex, Claude Code, or the shared architecture underneath both. This is the heart of Part 2: separating product decisions from agent design .

 Takeaway: most of what makes these systems agents is shared. What differs is implementation: models, tool schemas, permission systems, context strategies, UX. Same architecture — different product decisions.

 Discussion

 Separate the model from the harness

 Discussion question If we swapped the model but kept the entire agent harness, how much of the product would still remain?

 Use this question to separate model capability from agent-system design . In Part 3, you will keep the design and throw away everything coding-specific.

 What a strong answer covers 
 Most of the product remains: the tools, permission system, state, instruction files, UX, and the loop itself. What changes is reasoning quality. That is why “which agent is best” is as much a harness question as a model question — and why vendors can swap models underneath without renaming the product.