I ran the same structured interview on two AI coding agents — same repo, same questions, same sequence. No coaching, no editing. Each answer built on the previous one.
Agents tested:
- Claude Code (Opus 4.7, 1M context window, Claude App UI)
- Codex (GPT-5.5, ~258K context window, Codex CLI)
What the interviews surfaced:
- Context fills up in ways you don't control and can't fully see
- Irrelevant material is never free — even when it doesn't hurt quality, it costs tokens
- Task switching has a cost that no system tracks or warns you about
- Output quality degrades gradually, continuously, and invisibly
- Both agents said: "If I'm drifting, I usually don't know it." Almost word for word
protocol/— The 11 questions. Ask them sequentially in a fresh session, in any repo your agent can see. Don't reorder, don't paraphrase, don't preview what's coming.transcripts/— Full Q&A transcripts from both agents, cleaned for readability.
Pick any coding agent. Open a fresh session in a real repo. Ask the questions in order. See what your agent tells you.
If you run the protocol on a different agent or model, open an issue or PR with the transcript — comparative data makes this more useful for everyone.
- Part 1: I Interviewed Claude Code and Codex. Here's What They Told Me. — LinkedIn post with carousel summary
- Part 2: Coming this week in the Context Engineering newsletter
- Gate Calibration paper: arXiv 7494978 — the framework both agents independently derived in Q10
Transcripts are published verbatim. The interview protocol is free to use and adapt. Attribution appreciated.
Interviews conducted May 18-19, 2026 by Borja Odriozola Schick.