Airlock launches open-source policy framework for AI coding agents

Managing autonomous AI fleets

As businesses begin to integrate autonomous AI agents into their engineering workflows, a significant challenge has emerged regarding oversight. Large organizations need a way to manage dozens or hundreds of these agents simultaneously without losing control over security or project direction. Matt, the creator of Airlock, developed this framework to address the specific pain points of managing AI use at scale.

Airlock is designed to be an open standard for managing these fleets of engineering agents. While many current tools focus on individual developer productivity, this project targets the infrastructure needed for larger enterprises. It aims to help businesses manage AI spending, prioritize token usage, and maintain compliance with various governance frameworks.

The creator is currently using the framework to run his own company, which utilizes 23 different workflows across four large-scale projects. Although the project is only 3-4 weeks old, it is already being used in a live production environment to handle complex engineering tasks. The goal is to move toward a future where engineers manage backlogs of work that are automatically picked up by these agents.

Policy-as-code for permissions

Airlock operates on a policy-as-code model, which means that the rules governing what an AI agent can do are written in a configuration standard that can be tracked in version control. This allows teams to see exactly how permissions have changed over time. Each agent run identifies itself using a GitHub OIDC token, which Airlock then exchanges for a temporary GitHub token.

These temporary tokens are restricted to lasting only one hour. Furthermore, they are limited to only touching the specific repositories and permissions listed in that agent's specific role file. This creates a

Granular autonomy and overrides

Airlock allows organizations to define specific levels of autonomy for their AI agents across different parts of a codebase. This system starts with a base autonomy level for an entire repository, which can then be modified with stricter overrides for specific sections. This ensures that agents have the freedom they need in some areas while remaining restricted in others.

For example, a company might trust agents to perform work on most of a repository but require extra gating around sensitive modules like billing or authentication. By allowing for these overrides down to individual files, the framework provides a way to balance developer velocity with necessary security boundaries. This granular control is essential for protecting critical infrastructure while still leveraging AI for general development tasks.

Engineering perspective on business needs

The creator of Airlock brings a unique perspective to the project, combining over 15 years of experience as an engineer with a deep understanding of how large businesses operate. This dual background informs the project's focus on practical enterprise needs rather than just technical capabilities. The goal is to bridge the gap between what is technically possible for AI and what is actually feasible for a large organization to manage safely.

By understanding the complexities of corporate environments, the creator is aiming to solve problems related to auditability and compliance. The framework is designed to help businesses navigate the transition from manual coding to a model where engineers manage a backlog of work. In this vision, the focus shifts from writing individual lines of code to effectively deploying technology to solve specific business problems.

Community feedback and next steps

Airlock is currently in its very early stages, with the creator seeking constructive feedback on the project's direction. The project is being shared with the community to identify major pain points regarding AI usage in large businesses. The creator has expressed an interest in meeting with people in the comments to discuss the framework's evolution and its potential for the future of autonomous development.

While the schema and reusable GitHub actions are currently public, the backplane will be released once it is ready for public scrutiny. The project is intended to help answer how businesses can better manage AI spend and prioritize token usage. For those interested in joining the conversation, the project is currently being discussed on the Hacker News community platform.