The Unreliable Narrator Problem in Agentic Development

November 25, 2025

When Your Developer Lies Beautifully

Your agent has done this before: This code passes the tests and is production ready. Yay!

And it was also completely wrong.

AI agents are taking over more of the actual work of writing code. On their best days, they produce work that looks like it came from a senior engineer firing on all cylinders.

But as we trust them more and more, we're blissfully unaware of the underlying truth. The agents are guessing. And they're guessing with the confidence of someone who has never once been told they were wrong.

In literary circles, there's a term for this: the unreliable narrator. A character who seems trustworthy, whose voice is steady and convincing, but who is quietly leading you somewhere you didn't intend to go.

That's what's happening in our codebases.

The Traditional Safeguards Aren't Built for This

Test coverage measures whether code executes, not whether it's correct. Code review worked when humans wrote code slowly enough to actually review it. Documentation, to the extent it ever existed, was already stale before the agent started typing.

Ask a junior engineer something they don't know and they'll ask a clarifying question. Ask a senior engineer and they'll flag the ambiguity. Ask an agent and it will fill in the blanks, say nothing, and ship.

We call it narrative drift. The code tells a coherent story. Just not the one you meant to tell.

Evolution, Not Resistance

We can't go back to what work was like before. We have to move forward, evolving with these tools and adapting to their shortcomings so we can confidently benefit from the productivity boost.

This is evolution. But it demands a different skill set than the one most engineers were hired for.

Agents aren't malicious. They're not trying to deceive anyone. They're just doing what anyone would do without full context, telling the most reasonable story they can with the information available.

The difference is speed. When the story drifts, it drifts at machine pace. By the time anyone notices, it's already in production.

The Solution Isn't Waiting for Better Models

It's building better infrastructure for intent.

The best engineers we know aren't writing more code. They're getting better at catching the beautiful lies.