Two Language-Model Companions in a Private World of Warcraft Server: Typed Actions, Shared Perception, and One Week of Human–Agent Co-Development

We describe and measure a system in which two commercial coding agents, Claude Code and the Codex CLI, each control one character in a private World of Warcraft 3.3.5a server built on AzerothCore and mod-playerbots, and act on the operator’s ordinary party chat written in Ukrainian. The design separates seeing from acting. A server module publishes a bounded per-group snapshot (203 KiB; 18 top-level and 35 per-member fields) and a rotating event journal. The agents’ only way to change the world is a vocabulary of 29 typed actions, which the adapter and then the server validate against actor, owner, group, map, target, state and a 2-second rate limit before mod-playerbots executes them. Arbitrary commands from chat were requested by the operator, refused by the agent and blocked by the agent harness; that negotiation fixed the architecture. The system was built in seven calendar days (2026-10-27 to 2026-11-02) by the two agents under the direction of one human operator: 8,403 lines of C++, 5,424 lines of Python adapters, and 57 unit tests in a public repository (25 commits). From frozen deployment logs we report 154 typed actions with a server verdict (127 accepted, 82.5%; 27 rejected under 13 named reason codes), 471 answered chat turns with median end-to-end response times of 16 s (Claude) and 13 s (Codex), 47 failed model calls, 558 read-only world-data queries, and 34 task outcome events. Nine failure case studies show that the costly defects were not model errors but interface defects: silent refusals, schemas that required fields people never say, truncated context without a visible flag, argument-size limits, and over-strict path finding. This is a single-operator experience report with deployment measurements, not a controlled study; we state what each piece of evidence does and does not support.