AgentLoop is an open-source, zero-dependency Node.js developer tool designed to solve the context-degradation problem typical of long-running AI agent chats. Instead of maintaining a continuous session, AgentLoop starts a fresh Codex worker and critic every single cycle.

Users define their goal and establish a rubric using a GUIDELINES.md file just once. During execution, workers handle the building process while critics test the results. If failures occur, they are transformed into concrete fix notes for the next clean context, allowing project files to carry the necessary memory rather than bloated chat histories.

Runs are designed to stay local, sandboxed, observable, and fully cancellable via a live dashboard. Additionally, developers can leverage ChatGPT control through MCP. For higher standards, a special Polish mode can continue looping past a standard 'PASS' state until the critic formally gives a 'SHIP' status.

Pros of AgentLoop

Starts a fresh Codex worker and critic every cycle to prevent context degradation
Project files carry the memory through concrete fix notes rather than long chat histories
Runs remain local, sandboxed, observable, and cancellable from a live dashboard
Open source and built with zero-dependency Node.js
Supports ChatGPT control through MCP and features a Polish mode until the critic says SHIP

Cons of AgentLoop

Requires managing local execution and environment setup
Dependent on Codex workers and critics for task completion and testing
May require fine-tuning of the GUIDELINES.md rubric for optimal performance

Frequently Asked Questions

What is AgentLoop?

AgentLoop is an open-source, zero-dependency Node.js developer tool that starts a fresh Codex worker and critic every cycle, avoiding the memory degradation issues of long-running agent chats.

How does AgentLoop manage memory between cycles?

Instead of relying on a long chat history, project files carry the memory. Failures from the critic are turned into concrete fix notes for the next clean context.