BabyAGI

Breaks down high-level objectives into discrete subtasks using an LLM, then prioritizes and orders them based on dependencies and importance. The system maintains a task list in memory, executes tasks sequentially, and uses LLM reasoning to determine which tasks should be executed next based on completion status and goal relevance. This creates a self-directed workflow where the AI agent autonomously decides task ordering without explicit human choreography.

Executes individual tasks by passing them to an LLM along with the current task list, completed task results, and objective context. The LLM receives the full execution context (what's been done, what remains) and generates task-specific outputs. This allows the LLM to make decisions informed by prior work and avoid redundant or conflicting actions. Execution results are captured and stored back into the task list for subsequent tasks to reference.

Generates new tasks dynamically based on an initial objective and the current state of completed tasks. The system prompts an LLM to create the next set of tasks needed to progress toward the goal, using the objective and task history as input. This allows the agent to adapt its task list as it learns what's actually needed, rather than pre-planning all tasks upfront. New tasks are appended to the task list and prioritized for execution.

Maintains task state in a simple in-memory list structure (typically a Python list or JSON array) that tracks task descriptions, completion status, and results. The system reads from and writes to this list throughout execution, using it as the single source of truth for what's been done and what remains. State is not persisted to disk by default, existing only during the current execution session. This provides a minimal but functional state management layer without requiring a database.

Delegates task execution to an LLM by constructing a prompt that includes the task description, objective, and execution context, then parsing the LLM's text response as the task result. The LLM is responsible for reasoning about how to accomplish the task and generating an appropriate output. This approach treats the LLM as a general-purpose executor capable of handling diverse task types without task-specific logic. The system does not validate or structure the LLM's output; it accepts whatever the model generates.

Implements a main execution loop that repeatedly generates tasks, executes them, captures results, and generates new tasks based on progress toward the objective. The loop continues until a stopping condition is met (manual termination, max iterations, or objective completion). Each iteration uses the current task list and results to inform the next task generation, creating a feedback loop where the agent's understanding of what's needed evolves. This architecture enables the agent to adapt its strategy as it learns.