Part 2 of The Intellyx Building Agentic Resilience Series by Eric Newcomer, for Gremlin
The first part of this series identifies the core issue: generating application code using AI tools is not delivering promised productivity gains, mainly due to unresolved issues such as testing, and ensuring AI agents behave as expected.
One of the biggest challenges in deploying and governing AI generated code is discovering and preventing incidents and outages resulting from agentic hallucinations and errors.
It’s difficult to guard against coding agent errors that have unintended consequences, which can even be contrary to their explicit guidance.
Coding agents don’t have the big picture when they generate code, they only have the section of an application they are working on. When an agent finds and fixes a bug, for example, other parts of the code may be relying on the buggy behavior. If an agent decides to introduce multiple new logging points, it could degrade performance to an unacceptable level.
We are all witnessing examples of this challenge almost every day now as the AI coding agents grow more powerful.
Image source: Jan van der Wolf on Pexels.


