Why I Don't Trust an AI Agent to Have the Final Say in Development

Developers are increasingly turning to AI agents to automate coding. But should they be allowed to run in production unsupervised? One engineer decided not to, and shared his approach.

The Idea: Process, Not Freedom

The author isn’t trying to turn the agent into an independent developer. Instead, he proposes a strict framework where each step produces an artifact, and a human approves the transition to the next stage. The main steps are:

• SPEC — creating a specification
• PLAN — implementation plan
• TEST — writing tests
• CODE — code generation
• REVIEW — review
• LEARN — learning lessons

Key rule: at points where responsibility arises (for example, making a decision about deployment), the final decision rests with a person.

Why This Is Important

• It reduces the risk of uncontrolled changes to the codebase.
• The artifacts from each step allow you to trace the agent’s logic.
• The process remains predictable and reproducible.

The author calls his system map-framework and promises a series of articles: on hooks, contracts, context, and memory—everything he’s adapted from research papers into his workflow.

Conclusion

If you, too, want to use AI in coding but are afraid of letting the genie out of the bottle—it might be worth trying an approach that involves manually validating key decisions. The agent makes suggestions—you evaluate them.

Source: Habr