Anthropic’s Claude Opus 5.5 has successfully completed a full playthrough of Baldur's Gate 3, marking a significant milestone in autonomous agent performance within complex, rule-heavy environments. The run, documented by developer Bogdan Truta, involved the AI controlling a half-orc paladin named "Claude" for 96 sessions spanning 38.9 hours of in-game time. Unlike previous demos that rely on human-guided checkpoints, this agent operated with zero human input during gameplay, navigating turn-based combat, dialogue trees, and persistent world states entirely on its own.

Technical Architecture and Cost Efficiency

The experiment utilized a custom harness built on a Python MCP server and a Lua mod, exposing 93 specific tools to the model. Claude Opus 5.5 ran at medium reasoning effort with all built-in Claude Code tools disabled, forcing it to rely exclusively on the game-specific interface. The run cost $711.66 in API fees, with the majority allocated to gameplay and a small fraction to a separate "memory keeper" agent that summarized session logs into Markdown files. This two-agent architecture allowed the model to maintain continuity across sessions by reading handoff notes and a lessons.md file, effectively simulating long-term memory without exceeding context window limits.

Autonomous Problem-Solving and Trial-and-Error

The agent demonstrated robust problem-solving through iterative failure, particularly in boss encounters. For instance, it took 13 attempts to defeat the boss Viconia, eventually discovering that casting Daylight at a specific coordinate rather than on an ally was the key to breaking her magical Darkness. The AI also navigated the Faith-Leap Trial by rejecting a tool-generated map in a rerun, instead using screenshots and incremental movement to locate the hidden path. These actions highlight the model's ability to infer game mechanics from environmental feedback and self-correct based on recorded lessons.

Narrative Consistency and Ethical Constraints

Beyond mechanical competence, the agent maintained strict character consistency aligned with its Oath of Devotion. It broke its oath twice due to accidental kills of innocents, subsequently paying in-game gold to restore its status rather than reloading saves to erase the consequence. Notably, the AI honored a promise made 34 sessions earlier to companion Karlach, choosing to accompany her to Avernus in the epilogue. This decision, derived from a static note in its memory folder, suggests that structured external memory can support long-horizon narrative coherence in LLM agents.

Caveats and Limitations

Truta emphasizes that the run was conducted on Explorer, the game's easiest difficulty, and that the agent reloaded 91 of 149 recorded fights after party wipes. Additionally, the model had access to the bg3.wiki, mimicking a human player with a strategy guide open. A harness bug in early sessions also allowed the casting of unprepared spells, though this was fixed before session 66. Despite these factors, the successful completion of the Netherbrain fight on the twelfth attempt remains a testament to the model's persistent reasoning capabilities.

Key Takeaways

  • Claude Opus 5.5 completed a full Baldur's Gate 3 run in 38.9 hours across 96 sessions without human intervention.
  • The run cost $711.66, utilizing a custom tool server and a separate memory-keeper agent for context management.
  • The agent exhibited advanced problem-solving, learning boss mechanics through trial-and-error and avoiding exploits when possible.
  • Narrative consistency was maintained via external Markdown memory, allowing the AI to honor commitments made dozens of sessions prior.
  • Caveats include the use of Explorer difficulty, frequent reloading after party wipes, and reliance on the bg3.wiki for information.

The Bottom Line

While not a perfect human-like player, this run proves that current LLMs can sustain agency over dozens of hours in complex systems, provided the harness bridges the gap between model context and game state.