Why AI coding agents keep writing deprecated code (and what actually reduces it)
AI coding agents write deprecated code because they favor patterns common in training data, which lags current libraries. Here is why, and how to reduce it.
Blog
Notes on agent selection.
How coding agents choose and use developer tools, what the Fold Score measures, and what shape means for your SDK.
Agentic evaluation should be experimental science, not leaderboard competition
An agent benchmark score does not by itself tell you the agent will behave that way again, unless the choices behind it were fixed before the results appeared.
Accuracy is purchasable. Reliability is not.
A benchmark score climbs when you buy more retries. Repeatable behavior does not. Why agent accuracy and reliability differ, and must be measured separately.
Why your analytics cannot see agent tool selection
Your analytics can see an SDK was installed, not that an agent selected it. Why agent tool selection is invisible to product analytics, and how it is measured.
Agent evaluation variance: why a single agent run can be misleading
Agent evaluation variance is real: across 60,000 trajectories, ten identical benchmark repetitions spanned 2.2 to 6.0 percentage points per configuration.
GEO for developer tools in 2026: mentions are not market share
GEO for developer tools in 2026 is not brand mentions in ChatGPT but which SDK coding agents actually install. Why agent selection is the metric that matters.