Models

From Chat to Dark Factory: Six Levels of AI in Development

November 16, 2026 · 2 min read

a star filled sky
Aedrian Salazar / Unsplash

For over a year, AI has been integrated into the development of a mature SaaS company, alongside an established engineering team that was initially skeptical about AI. The team has progressed from copying code from an AI chat to allowing agents to handle feature development entirely. This approach is described as AI engineering, focusing on building a system that consistently meets quality standards. Currently, the team is experimenting with a semi-autonomous agent factory in the cloud.

The journey has been categorized into six levels, from traditional manual development to a "dark factory" model. Each level represents a shift in the development lifecycle, detailing the role of humans and how teams can assess their progress and what is necessary to advance further.

At L0, manual development is the baseline where each lifecycle stage is performed by humans, from gathering requirements and coding to releasing and monitoring. Two years ago, requests for new features and support were managed manually, and project specifications existed only in the heads of developers. Some services operated on trunk-based development, with commits directly to the master branch without pull requests or reviews. Quality assurance was performed manually according to outdated project specifications, though some projects had good test coverage and static analysis set up, which is crucial for agent development.

L1 introduces an assistant level where individuals consult with an AI chat tool for advice on specifications, design, and code improvements. Users can copy code snippets and generate test cases with the help of the AI, making the development process faster without changing its essence. The organization needs to provide access to approved AI tools and clarify their usage.

At L2, teams work with an agent that is being trained according to team rules. Here, quality gates, pull requests, and context files are essential to ensure that the agent learns effectively without overwhelming itself with information. The agent is granted access to necessary tools and internal data, enabling it to verify its work and read the same documents as human team members. Clear documentation of recurring issues helps prevent the agent from making the same mistakes repeatedly, while training standard procedures into skills ensures adherence to team agreements.