Today, AI can generate a form, a workflow, or even part of an application in minutes. The real challenge begins when those changes are deployed to a system that users rely on every day. How can you leverage AI agents while maintaining full control, ensuring security, and preserving the ability to review, approve, or roll back changes when needed?
An AI agent can significantly accelerate the creation of new features and the development of applications. However, when it starts working on a production system that users rely on every day, a new challenge arises: ensuring the security of those changes.
A system that supports business processes must operate in a predictable and stable manner. Whether it handles overtime calculations, sales monitoring, or equipment order approvals, every change should be controlled, verifiable, and fully auditable.
Coding agents are excellent at prototyping, iteratively developing functionality, and responding to user feedback. In practice, however, a single prompt is rarely sufficient. The final solution is most often the result of multiple iterations between the user and the AI model. To fully leverage the agent’s capabilities, the environment in which it operates must be designed specifically for this approach to software development.
Contact us, and we’ll be happy to demonstrate what the system can do.
You could say that we at Productive24AI have been fortunate, although that “luck” was the result of deliberate architectural decisions. Even before AI became widely adopted, we decided that GitOps would be the best approach for managing our system configuration.
Git is the standard version control system used in software development. GitOps is a methodology that uses a Git repository as the single source of truth for a system’s state. It enables teams to track the history of changes, allows multiple people to work in parallel, and provides full control over the application development process. This approach offers another important advantage: Git is already a familiar environment for developers. There is no need to teach Claude Code or Codex how to work with a repository – it is their natural working environment. When configuring Productive24AI, Git primarily ensures the security and complete auditability of all changes.
The configuration implementation process consists of three stages. Before discussing them, however, it is worth explaining the concept of a “branch.” A branch is an independent version of a repository created to develop new features or introduce changes to existing solutions. It allows you to work on the configuration without affecting other system users. We recommend creating a separate branch for each feature, as this makes it easier to manage changes and track their history later.
The Configuration Deployment Process in Productive24AI:
Working on your own branchEvery change begins with the creation of a new branch. It is good practice to give the branch a name that clearly describes the scope of the work being performed. Next, the user – either independently or with the support of an AI agent – prepares all necessary modifications.
Merging into the main branchOnce the work is complete, the changes are merged into the main branch, which serves as the testing (QA) environment. Here, changes prepared by different users can be verified together, making it easier to identify potential conflicts or unwanted dependencies.
Deployment to productionThe final step is merging into the release branch, which reflects the configuration of the production environment. Only at this stage are the approved and tested changes delivered to end users.
At any stage, it is possible to launch the configuration in test mode. This allows new features to be tested without affecting other system users. Additionally, selected changes can be tagged and made available exclusively to a specific group of testers.
The AI agent is provided with a complete set of tools required to work with the system – from creating and modifying configuration elements, through accessing documentation, to navigating and searching for the appropriate components. All operations are immediately available while ensuring that the system’s structure and consistency are maintained.
In addition to the tools provided through MCP servers, we also allow direct editing of files within the repository. Coding agents were designed to work with code, so we do not restrict this capability. Instead, we provide guidance on when it is more appropriate to use MCP tools and when to edit files directly, enabling more efficient token usage. However, the ability to edit files directly does not compromise security. As a low-code platform, we have built configuration validation mechanisms into the platform from the very beginning. Forms, lists, charts, and other elements are automatically checked for correctness.
If an agent makes a mistake or creates an incorrect configuration, the system detects it before deployment. Validation is performed continuously and before each merge stage. Error information is sent both to the user and – through MCP tools – to the AI agent, which can then correct the configuration independently. Only after successful validation can changes be deployed to the production environment.
For every merge operation, we provide a report showing all changes that will be introduced. Regardless of whether the changes were prepared by an AI agent or a user, you can thoroughly review which elements will be added, modified, or removed before deployment.
Despite the safeguards in place, there may still be situations where implemented changes need to be rolled back. Users may not yet be ready to adopt new functionality, or unexpected issues may occur after deployment. Thanks to Git, reverting to a previous version of the configuration is simple and can be done using the revert operation, allowing you to quickly restore a proven system state.
revert
The mechanisms described here are not new. Git has been the standard for version control for years, test environments are an integral part of the software development process, and validating changes is a standard step developers perform before releasing code. Our goal was not to create new rules for working with AI, but rather to apply proven software development best practices. This allows us to safely develop existing applications, leverage AI agents, and maintain full control over every change.
Our GitOps-based architecture made integration with Claude Code or Codex a natural next step rather than requiring a platform overhaul. This allowed us to focus on what matters most – teaching AI agents how to effectively support users in building applications.
AI agents significantly accelerate application development, but their use in production systems requires appropriate control mechanisms. GitOps, change validation, test environments, and a complete version history make it possible to leverage AI without compromising the security and predictability of deployments. It is precisely this combination of proven software development practices and the capabilities of AI agents that enables faster application development while maintaining full control over every change.
MVP in a couple of days!
Test Productive24AI in your business processes!
to our newsletter and stay up to date with new publications on the blog.
Try it out at no risk!
