Trial by Combat turns LLM benchmarking into duels

Hot take: this is less a classic benchmark and more an agent-performance stress test, which is exactly why it’s interesting.

• –The open-source setup makes the comparison reproducible, which is stronger than a one-off demo clip.
• –The match outcome suggests speed under low-reasoning settings can matter as much as raw model quality in turn-based agent tasks.
• –Hidden information, traps, and simultaneous turns are a good fit for evaluating planning, not just text generation.
• –The curl-native API lowers friction for running arbitrary model-vs-model duels, which is a neat systems design choice.
• –If the repo keeps matches deterministic and replays exact, it could become a useful sandbox for agent benchmarking and prompt iteration.