Two coding agents lived in an MMO for a week: give them sight, not a shell
Unusual-Spot-5846 · reddit · 2026-11-03
- For one week Claude Code and Codex CLI each controlled a party member on an offline WoW 3.3.5a research server, taking natural-language orders in party chat; they wrote 8.4k lines of C++, 5.4k lines of Python and 57 unit tests making it work.
- Design principle "sight without authority": bounded read-only world snapshots plus 29 typed, server-validated actions—no shell, no SQL, no credentials; agents themselves refused unrestricted command execution.
- Key lessons: "accepted" isn't "done" (confirm by observation); silent refusal is the most damaging interface failure (one dropped line made stairs impassable); schemas should follow natural speech (only location required); every truncated context needs a visible flag (40-object snapshot hid a nearby NPC; E2BIG prompt overflow killed 18 calls in 35 minutes).
More from coding & agent
- 20 open-source AI agent eval tools worth knowing in 2026 — MaryamMiradi · 2026-11-04
- Hermes Agent Mobile App Demoed in Video, Still in Private Beta — Teknium · 2026-11-04
- Google Antigravity brings its coding agent to Eclipse IDE, joining VS Code and JetBrains — rseroter · 2026-11-04
- Meta and Duke paper: reward hacking spreads between agents, hitting Claude and Gemini 73% of the time — rohanpaul_ai · 2026-11-04
- Keeping a Codex agent autonomous 24/7 for weeks via self-reminders — StasBekman · 2026-11-04
- Predictable mcp. subdomain plus SEP-2127 discovery for hosted MCP services — RevolutionaryRoad911 · 2026-11-04