deepfates/cantrip

Cantrip is a system designed to transform large language models (LLMs) from stateless predictors into stateful actors capable of interacting with their environment through code execution and tool use. It introduces a runtime where LLMs operate in self-modifying loops, allowing entities—emergent agents defined by the system—to iteratively generate, execute, and refine code based on observations from their environment. This approach enables more complex, adaptive behaviors than static, pre-defined agent architectures.

The core architecture of Cantrip centers on the concept of a "Circle," which defines a bounded operational space for each entity. Within this space, an entity is shaped by three primary components: the LLM itself as the cognitive engine, an immutable Identity (including system prompts and model hyperparameters), and the Circle, which comprises a Medium (such as conversation, code, or Bash), Gates (interfaces for external interaction like file reading or calling other entities), and Wards (runtime constraints such as maximum turns or sandboxing rules). The entity's action space is formally defined as the union of available mediums and gates, minus any imposed wards.

Cantrip supports two principal modes of agent instantiation: "Cast," which creates a one-shot entity for a single intent that terminates upon task completion or when a constraint is met, and "Summon," which spawns a persistent, supervised process that maintains state across multiple interactions. The state and history of each entity are recorded in the "Loom," an append-only log that can be backed by storage systems like Mnesia or JSONL, allowing for durable memory and the ability to resume context across sessions.

The primary runtime for Cantrip is implemented in Elixir/OTP, leveraging components such as GenServer for state management, DynamicSupervisor for process lifecycle control, and sandboxed execution environments for code. The system also includes a pre-configured coordinator called the Familiar, designed to operate within a codebase with workspace observation capabilities and durable storage. Cantrip is positioned as a framework for building advanced, adaptive agents that can analyze, modify, and interact with environments through iterative language-to-action cycles.