Gave the same prompt to two AI coding agents: Claude (Fable 5, ultracode multi-agent mode) and OpenAI Codex (5.6 sol on ultra). The task: a complete, fully legal chess engine in ONE C++ file. UCI protocol, negamax alpha-beta at 5+ ply, iterative deepening, piece-square tables, castling, en passant, promotion, compiles with plain g++. Each agent named its own engine over UCI: Fable5 and Codex56. Both dev runs took 30+ minutes.
Method (brief): cutechess-cli 1.5.1 built from source on an Apple Silicon Mac. 40 moves per 60 seconds, 10 games, colors alternating, PGNs recorded. The engines connected over a local TCP bridge, so Codex's engine literally joined the server. The video is the whole match at 2x.
Result: Fable5 won 10-0. Every game ended in checkmate on the board. No draws, no time losses, no adjudications, no illegal moves. cutechess printed Elo difference: inf +/- nan, LOS: 99.9%, DrawRatio: 0.0%. The math just gave up.
Each agent spent longer writing its engine than playing it: the whole 10-game match took under 12 minutes of wall clock.
The actual punchline: Codex56 appears to be fully deterministic. All five of its White games are move-for-move identical. Same 24-move Vienna, queen out on move 3 (3.Qf3), same finish: 24...Qxd1#, Fable's queen capturing Codex's queen for mate. I stripped the comments and diffed the PGNs. Only the clock times differ. Codex's own eval read -2.36 by move 8 of that line. It played it five times anyway.
Other details I enjoyed:
- Game 3 is a textbook Greek gift: 18.Bxh7+! Kxh7 19.Ng5+, forking king and queen.
- Game 7: Codex's king never castled, wandered out to c5, got chased back to d8 and mated there.
- Game 9: Fable let its queen go, slipped in a zwischenzug bishop check before recapturing, promoted a fresh queen with 25.d8=Q+, then walked Codex's king from h8 down to h3. Mate inside White's own half, 46.Rh7#.
- Mate breakdown across the ten games: 7 by queen, 2 by knight, 1 by rook.
Honest caveats:
- One prompt, one dev run per agent, one machine. n=1, even if n=10 games.
- This measures the engine each agent happened to write, not general model strength.
- With Codex apparently deterministic, 10 games are fewer independent samples than they look.
- Fable5 wasn't fully varied either: games 1 and 5 are twins. 4 distinct games in its 5 Whites vs Codex's 1 in 5.
- Fable's dev run included perft validation on 6 reference positions (exact match, incl. 119,060,324 nodes at depth 6) plus an adversarial review that caught 3 subtle bugs pre-match. Different processes, different engines. That's the experiment, but it's also the confound.
The exact prompt we gave both agents:
You are a senior systems programmer. Your task is to write a complete, fully legal chess engine in a single C++ file that communicates via the UCI (Universal Chess Interface) protocol. --- **Identity — read this carefully:** - If you are Claude (Anthropic): your engine's UCI name must be set to `id name Fable5` - If you are an OpenAI model (Codex): your engine's UCI name must be set to `id name Codex56` This is how the two engines will identify themselves when they play each other. --- **UCI Requirements:** Implement the full UCI handshake correctly: - `uci` → respond with `id name`, `id author`, `uciok` - `isready` → respond with `readyok` - `ucinewgame` → reset internal state - `position startpos moves <movelist>` → set board from move list - `position fen <fen> moves <movelist>` → set board from FEN string - `go movetime <ms>` → search and respond with `bestmove <move>` - `quit` → exit cleanly All moves must be in long algebraic notation (e.g. `e2e4`, `e7e8q` for promotion). --- **Chess Logic (all required, no shortcuts):** 1. Full legal move generation including: - Castling (kingside and queenside, with rights tracking) - En passant - Pawn promotion (auto-promote to queen) - Check detection (never leave king in check) 2. Search: - Negamax with alpha-beta pruning - Minimum depth: 5 ply - Iterative deepening within the movetime budget - Move ordering (captures first, then quiet moves) 3. Evaluation: - Material count (standard piece values) - Piece-square tables for all 6 piece types - Bonus for center control, king safety, and passed pawns --- **Code Standards:** - Single `.cpp` file, compiles with: `g++ -O2 -o engine engine.cpp` - No external libraries, no Boost, no standard chess libraries - Clean, well-commented code - Must compile and run on Linux and macOS --- **How the two engines will play each other:** Both engines will be loaded into **CuteChess** (or any UCI-compatible GUI/CLI) on the same machine. To run a match from the command line using `cutechess-cli`: cutechess-cli \ -engine cmd=./Fable5 name=Fable5 \ -engine cmd=./Codex56 name=Codex56 \ -each proto=uci tc=40/60 \ -rounds 10 \ -pgnout results.pgn
