Don't rewrite your CLI for AI agents

In recent months, there has been a growing recommendation to redesign command-line interfaces to exclusively accept JSON payloads, based on the premise that artificial intelligence agents work better with hierarchical structures. However, a recent study with language models of different sizes shows that this intuition does not hold up when measuring real efficiency. Traditional flat arguments (flags like --name value) not only match JSON in accuracy for the most powerful models, but clearly outperform lighter ones, where JSON causes syntax errors, shell escape errors, and token consumption up to fourteen times higher. This finding is crucial for any company wanting to integrate AI agents into their workflows: it is not about making expression easier for the model, but about limiting its error space. Tools that work with humans usually also work with models, because they restrict valid options and eliminate ambiguities. At Q2BSTUDIO, as experts in artificial intelligence and custom software development, we constantly observe how design decisions based on assumptions —no matter how plausible they seem— must be contrasted with real data. Building interfaces for AI agents is no exception to this rule: offering a --json flag may make sense for programmatic integrations or batches, but eliminating flat arguments entirely introduces a model dependency that not all systems can assume. Furthermore, token costs skyrocket when the agent has to correct shell escapes or redo nested structures. This behavior is independent of the cloud provider: in both AWS and Azure environments, legacy CLIs maintain their effectiveness. Therefore, instead of hastily rewriting a functional interface, we recommend first evaluating with real scenarios, measuring accuracy, cost, and execution environment. At Q2BSTUDIO we apply this philosophy also to our cybersecurity services, business intelligence services, process automation, and custom application development, where empirical evidence guides every architecture. The AI agents we build for our clients benefit from CLIs with well-documented flat arguments, reducing friction and operational costs. Even in combination with Power BI dashboards or monitoring systems, maintaining the simplicity of the underlying interface ensures that smaller models —the most economical— remain viable. The final reflection is clear: before getting carried away by technical trends, it is advisable to measure. A one-day experiment with the models and shells your team actually uses can save months of rewriting and unnecessary costs. In practice, the best interface for an AI agent is not always the most expressive, but the one that best adapts to the constraints the model can reliably handle.