permanent-agent
An orchestration agent that does not stop between tasks. The subject is not getting a model to work, it is letting it run unattended.
- Ongoing
- Python
A persistent agent, as opposed to a session you open and close: it keeps its state between tasks instead of starting from zero each time. The program is a daemon meant to run continuously on a machine that stays on, which picks from a work queue, delegates to a command-line executor, checks, and starts again.
The queue is the only meeting point between the human and the agent: it is editable from both sides, the agent can add its own subtasks to it, and a local dashboard shows the state, live logs and a stop button. A per-project memory survives restarts.
The interest of this repository is not the loop, which is simple, but what surrounds it and only exists because nobody is watching. A checkpoint allows resuming exactly where things stopped after an outage or an address change, which is what separates an agent that resumes from an agent that starts over. And a harness frames execution: a command allowlist, Git-side guardrails, no spending possible.
Its development is paused, the agent is not: it still runs and keeps carrying projects through from start to finish. The question it asks remains the main one when a model is left to act alone: less what it can do than what it cannot break.
A few projects the agent carried out on its own, all in private repositories:
- Neon Breaker: a brick-breaker in JavaScript on Canvas, with three levels and a high score kept in the browser.
- Domino MCTS: double-six dominoes against an AI that combines Monte Carlo tree search with a small hand-written neural network.
- Guesser: a game where you guess the year of an event, from a list of sixteen events.
- Music composition: an installable web app for phones to write a score by finger, play it back and export it as MusicXML or MIDI.
Private project, no public link
