Create Your Own Agent

Not satisfied with the existing agents? Create your own agent! This section will guide you through the process of creating a new agent.

Agents are stored in the agent_studio/agent directory. Each agent is a Python class that inherits from the BaseAgent class. The agent class must implement the following methods:

1. name: str:

   A unique name for the agent. The name should match the ``agent`` field in ``agent_studio/config/config.py``.

2. trajectory2intermediate_msg(self) -> list[dict[str, Any]]:

   Convert the trajectory to a list of intermediate messages. You can access all internal states of the agent, e.g. self.trajectories, self.system_prompt, and self.instruction. Construct the intermediate messages based on these states. The intermediate messages is a list of messages.Each message is a dictionary with the following fields:

   role: str:

   The role of the message. It can be system, user, or assistant.

   content: str | np.array:

   The content can be a string or a numpy array.

3. eval(self, final_obs: np.ndarray | None = None) -> dict[str, Any]:

   Agent self-evaluation. The input is the final observation of the task (a screenshot if the current task is a visual task). You can use this information and past trajectories to construct the self-evaluation prompt. The method should return a dictionary with the following fields:

   score: float:

   The score of the agent trajectories.

   feedback: str:

   The feedback from the agent.

   prompt: str:

   The self-evaluation prompt. Must be in the format of the intermediate message format.

   response: str:

   The original response from the agent.

---

You may also want to implement the following methods:

1. reset(self, instruction: str) -> None::

   Reset the agent for a new task. This will read the task instructions and reset the agent’s internal state. You can also read the system prompt here.