Knowledge cutoff

The knowledge cutoff is the point in time after which a model has seen nothing. Training data is collected up to a date, the model is trained, tested and released some months later, and then used for a year or more. So at any moment the model's picture of the world is stale by somewhere between several months and a couple of years, and nothing in the harness moves that date.

For coding this bites in a specific way: libraries move faster than models. A package released after the cutoff, a breaking major version, a renamed config option, a new CLI flag; the model has no parametric knowledge of any of it, and by default it will not say so. It will write the old API with total confidence, and the resulting error will look like your mistake rather than its blind spot.

Assume the cutoff is a problem whenever the thing you are using is newer than about a year old. Load the current docs or the installed package's own type definitions into the context window, and tell the agent explicitly which version you are on. A sentence in your project file ('Next 15, app router, Tailwind v4') prevents a whole category of confident regressions.