I made an observation recently about how coding agents write code and maybe the community has some feedback.
To me, it seems that agents write code that “just doesn’t want to fail”. This means swallowing errors or providing defaults that don’t make sense. This way, errors only surface far away from their origin.
For example, today I tasked Codex to pull a value from a Python dictionary, do something with it and then put it inside a pydantic model. It decided to get the dictionary value like this:
<pre><code>value = d.get(&quot;key&quot;, &quot;&quot;)
</code></pre>
Unfortunately, <code>&quot;key&quot;</code> was not in the dictionary, so every <code>value</code> was an empty string. Only much later an error was thrown because the pydantic model expected a non-empty string. If I had written the code by hand, I would have used:
<pre><code>value = d[&quot;key&quot;]
</code></pre>
This way, the program would have failed at the “correct” position. It’s a simple example, but I hope it brings my point across.
Has anyone had a similar experience?

Tilman Krokotsch

A developer observes that AI coding agents tend to write defensive code that suppresses errors rather than failing fast at the correct location. Using a concrete Python example with Codex, the author shows how an agent used `dict.get('key', '')` instead of `dict['key']`, causing a silent failure that only surfaced much later in a Pydantic model validation. The post raises the question of whether others have noticed this pattern of agents preferring silent defaults over explicit failures.

Coding agents don't want their code to fail

In a way it makes sense, just like when chatting they are overly agreeable, sycophantic, in code they want code that works no matter what so the user is happy and keeps using it