SageCraft
An open-source agent framework built around the Sage decision model, designed specifically for playing World of Warcraft via screenshots and keyboard input. The architecture separates concerns between a harness responsible for perception and execution and the model responsible for decision-making.
Timeline (last 90 days)
SageCraft is the first harness (agent framework) built around the Sage decision model, with the goal of playing World of Warcraft entirely through screenshots and keyboard input, without modifying game files or using plugins.
The entire SageCraft system runs on macOS and consists of three components: the Sage model, the harness framework, and simulated keyboard and mouse operations.
SageCraft's design philosophy is that the harness is responsible for perception and execution, the model is responsible for decision-making, and the harness provides the model with a list of available actions and validates the model's selections.
In the experiment, the dwarf character leveled up to level 5, consuming a total of 4,573 decisions across 32 game sessions, with a cumulative play time of 5 hours and 29 minutes.
The entire experiment was supervised by the developer, who manually fixed harness issues between each session, and was not a fully automated, unattended AI player.
The final 18 minutes of leveling from the late stage of level 4 to level 5 was completed without any human intervention.
The developer has made SageCraft's harness code publicly available and recorded a build video for other researchers to reproduce and improve upon.
All Facts (7)
SageCraft is the first harness (agent framework) built around the Sage decision model, with the goal of playing World of Warcraft entirely through screenshots and keyboard input, without modifying game files or using plugins.
50%UnverifiedThe entire SageCraft system runs on macOS and consists of three components: the Sage model, the harness framework, and simulated keyboard and mouse operations.
50%UnverifiedSageCraft's design philosophy is that the harness is responsible for perception and execution, the model is responsible for decision-making, and the harness provides the model with a list of available actions and validates the model's selections.
50%UnverifiedIn the experiment, the dwarf character leveled up to level 5, consuming a total of 4,573 decisions across 32 game sessions, with a cumulative play time of 5 hours and 29 minutes.
50%UnverifiedThe entire experiment was supervised by the developer, who manually fixed harness issues between each session, and was not a fully automated, unattended AI player.
50%UnverifiedThe final 18 minutes of leveling from the late stage of level 4 to level 5 was completed without any human intervention.
50%UnverifiedThe developer has made SageCraft's harness code publicly available and recorded a build video for other researchers to reproduce and improve upon.
50%