AgentLoop: Stop Being the Relay Between Planning and Code

Research Context and Background

AgentLoop is a local-first development tool designed to eliminate the manual relay between high-level planning and code execution. It allows developers to set a goal and standards once, after which the tool autonomously handles every handoff between the planner, the worker, and the critic. This system lets developers stay in control without managing every intermediate message. The tool runs within your local environment, requiring Node.js 18, Git, and an authenticated Codex CLI. It does not require npm install or a hosted workspace. The system operates on a continuous loop of planning, execution, and rigorous review until the project passes the defined standards.

Benefits

AgentLoop offers several key advantages for developers. First, it uses fresh workers for every cycle to ensure clean context and prevent context drift. Second, it employs an independent critic that reviews results separately. This dual approach ensures that while tests pass, the code also meets high-quality standards and edge cases are covered. Third, the tool is local-first and open source, which preserves data privacy and reduces dependency on external hosted workspaces. It is MIT licensed and designed with a pluggable engine boundary for future integrations. Finally, developers can observe every cycle, verdict, transcript, cost, and file handoff in real-time. The system can be cancelled from the dashboard or monitored via ChatGPT through the Model Context Protocol.

Use Cases

AgentLoop is intended for work that would otherwise require babysitting. Common use cases include implementing new features, database migrations, code hardening passes, and product polishing. For example, a developer might set a goal to ship a complete feature like adding CSV export across the UI, API, and regression suite. In the first cycle, the critic might fail the cycle because a fresh critic identified a real mixed percent-decoding defect that the existing suite missed. The system then generates a fix. In the next cycle, the worker applies the fix, adds regression coverage, and clears the rubric. Once all tests pass and the defect is resolved, the system returns a verdict of pass. This workflow allows developers to focus on the goal while the system handles the complex handoffs and quality assurance required to reach a shippable state.

Pricing

Pricing details are not available in the provided information.

Vibes

Public reception and specific reviews are not available in the provided information.

Additional Information

AgentLoop is open source and MIT licensed. It is designed to run locally without a hosted workspace. The tool supports future integrations such as more engines or research loops through its pluggable engine boundary. The system relies on a static definition of done stored in a GUIDELINES.md file. Prompts may disappear, but the standards remain. The tool was created to shift development from manual babysitting to automated building by separating planning from local execution and ensuring rigorous independent review.