Harness Engineering - Beyond the AI Model
AI can identify a failure and suggest a fix. Should we trust it to fix the problem on its own? Why AI-driven remediation needs production-grade discipline.
AI can identify a failure. It can even suggest a fix. But should we trust it to fix the problem on its own?
I’ve been thinking about this quite a bit while working with AI-enabled engineering and automation platforms.
Anyone who has spent enough time in software engineering knows that not every failure is what it first appears to be.
A test fails. The initial assumption is that something is broken.
Maybe the application is perfectly fine, and the test data has changed.
Maybe a downstream service is responding slowly.
Maybe an element isn’t loading in time. Or perhaps there’s a genuine defect that needs attention.
Now imagine an AI agent investigating that failure.
It looks at the logs, analyzes execution history, identifies a probable cause, and recommends a fix.
That’s useful. It can save engineers a significant amount of investigation time. But here’s where things get interesting.
What if the agent gets the diagnosis wrong?
What if it updates an automation script to make the test pass, while the actual application defect remains undetected?
We’ve solved the immediate failure, but potentially created a much bigger problem.
This is one of the reasons I believe AI-driven remediation needs to be designed with the same discipline we apply to production engineering.
An agent shouldn’t just recommend a fix.
It should be able to explain the evidence behind it, validate the change, and recognize when the decision needs human intervention.
And sometimes, the most valuable thing an AI agent can do is simply say, “I’m not confident enough to make this change”.
To me, that’s not a limitation of AI. That’s a sign of a well-engineered system.
The goal shouldn’t be to make every failed test turn green automatically.
The goal should be to make sure we’re fixing the right problem.
That’s the distinction I think will matter as we move toward more autonomous engineering platforms.
Originally shared on LinkedIn.