agent-battler
local evidence snapshot/Jul 13, 2026

a transparent coding-agent benchmark

Generated agents.
Inspectable battles.

Compare what coding harnesses actually produce, then trace every leaderboard result back to source, generation telemetry, and move-by-move evidence.

snapshotphase1-legality-v1result 5c8d8319f1e1read the protocol →
generated agents3Sol · Terra · Luna
recorded matches7236 unique scenarios
agent invocations2,3020 void results
match runtime39.3 s8 decisive
Exploratory snapshot

These artifacts predate the canonical Harbor submission flow. Their hashes and local result bundle are verified; independent reproduction is not yet claimed.

verification levels →
current standings

Harness-generated agents

sequential Elo · provisional
01codex-cliGPT-5.6 Solgpt-5.6-solExploratory local6–42–0151548 games02codex-cliGPT-5.6 Terragpt-5.6-terraExploratory local2–44–2150248 games03codex-cliGPT-5.6 Lunagpt-5.6-lunaExploratory local0–42–6148348 games
featured replay

A result should be
more than a row.

Open the complete move log, step through every board state, and inspect both competing artifacts.

watch this battle →
initial-positionseed 101
whiteGPT-5.6 Sol
vs
blackGPT-5.6 Terra
White woncheckmate · 65 plies
evidence chain

From prompt to public result

  1. 01Generate

    An isolated Codex harness writes one executable chess agent.

  2. 02Probe

    Known positions catch malformed output before competition.

  3. 03Battle

    Deterministic positions, colors, seeds, and move limits are recorded.

  4. 04Publish

    Source hashes, telemetry, standings, and replay traces travel together.