The reason AI agents shouldn’t touch your source code — and what they should do instead

Dynatrace is experiencing a major moment in its history. It’s expanding from an observability platform what might be described as a new environment, one that offers configurations for autonomous operations and security.

In this episode of The New Stack Makers, Alois Reitbauer, Chief Technology Strategist at Dynatrace, unpacks his predictions for a world where agentic AI manages production environments.

This interview, recorded at the Dynatrace Perform conference in late January, took place just a few weeks after Dynatrace acquired DevCycle, the feature-management platform. Feature flags have been around for decades, but with agentic AI, they may have a new role as something of a safety switch to stop runaway AI. In a way, it’s a solution that’s been hiding in plain sight.

DevCycle acts as the safety mechanism, in the words of Reitbauer, that enables operations to truly be autonomous.

“The safest change to an environment is not rewriting the code and deploying it by an agent, but changing the configuration,” Reitbauer tells The New Stack.

Here’s one memorable exchange from the interview:

The New Stack: There’s a lot of anxiety about agentic AI making decisions they shouldn’t and making consequential ones. Do you see feature flags as a safety mechanism?

Reitbauer: It is a safety mechanism. It’s not that we are assigning our production environments over to AI; that’s not where we are. But there are a lot of tasks where AI can make faster decisions and can also take faster actions. And that’s also the feedback that we’re actually getting from our customers. They say they actually find AI acting within certain guardrails and constraints.

“This might be workflows that we’re exposing via [model context protocol], and an agent can choose [a company’s] own books that the agent chooses, and feature flags perfectly fall into this category. You allow the agent to change some of those configuration settings, but you’re more or less options to pick and choose. It’s a very safe mechanism.

“What nobody would want is when an agent sees a problem in production and just decides on its own to totally reconfigure it. It’s more [about] making those proper choices. We also see that our customers have a way more appetite to move in that direction.”

It’s a fascinating future, and Reitbauer seems keenly tuned into how it will unfold.