Do coding agents show consistent tool selection bias?

<!-- SC_OFF --><div class="md"><p>An interesting observation: I’ve stopped using the LLM-powered autocomplete in my IDE. </p> <p>At first, it was one of the key features for me.

<!-- SC_OFF --><div class="md"><p>everyone's focused on the benchmarks (14 state of the art evals, beats mythos on terminal-bench, etc) but i think the actual story of 5.5 is what

<!-- SC_OFF --><div class="md"><p>Cursor dropped their new product yesterday and its not a code editor anymore. its an agent orchestration console that happens to have an editor

<!-- SC_OFF --><div class="md"><p>It’s been somewhere between 1 and 1.5 years since I last wrote a line of code.</p> <p>I wrote everything from Assembly and C to Python and

<!-- SC_OFF --><div class="md"><p>The productivity metrics on AI coding tools focus almost entirely on acceptance rate and time saved. The metric nobody is tracking is technical

<!-- SC_OFF --><div class="md"><p>I'll be upfront about my endgame: I work at a large tech company, I have a niche picked out, and I'm making the move to build and sell AI agents

<!-- SC_OFF --><div class="md"><p>I’m starting to think most “agent bugs” aren’t bugs. They’re mismatches between what we think we asked and what the agent thinks we asked.</p>

<!-- SC_OFF --><div class="md"><p>i’ve been playing around with coding agents recently and kept running into the same issue:</p> <p>they get stuck in loops</p> <pre><code>fail →