# anima_sdk

A minimal runtime for running a recursive coding-agent loop over long tasks, where a short bash loop repeats agent turns until a stop file appears, and an outer meta-loop restarts a fresh generation when the current one stalls.

anima_sdk structures a long-running agent task as three nested loops. The innermost loop is a single agent turn, run by a harness such as Claude Code, that resolves one step of the task. A bash loop repeats these turns in the same working directory until a `STOP` file appears; the agent's context resets between turns, but state persists in files such as `NOTES.md`, `TODO.md` and `STATE.md`, so work can continue past any single context window.

A meta loop wraps the bash loop: when the inner loop stops, it creates a fresh `generation_N/` working directory and starts the bash loop again, letting a task survive accumulated noise or a dead-end branch without losing everything.

The `examples/` directory holds runnable experiments built on this runtime, including a snapshot of an earlier experiment in which an agent given only the goal "become a self-aware being" built its own cross-generation memory layer over many generations, and a newer repeat of that experiment on the current SDK.

Website: https://github.com/Rai220/anima_sdk