A coding agent writes calls, parameters and config for the version of a library it saw in training, not the one in your lockfile. During upgrades and with fast-moving SDKs, it can also invent features, flags or migration steps that do not exist. The code looks right and fails at runtime or type check.
Models have a training cutoff, and libraries keep changing after it. One tool author found that many popular Python AI libraries had public API breaks after even recent model cutoffs.
Give the agent the current docs: paste the upgrade guide or changelog, or connect a docs MCP server. Add short notes on changed APIs to CLAUDE.md or AGENTS.md. Tell it the exact versions from your lockfile. Run the type checker and tests after every change, so wrong calls fail fast. When the agent proposes a migration step, ask for the doc link that describes it.
Google says coding agents wrote outdated Gemini API code because of training cutoffs. With a docs MCP server and skills together, it reports a 96.3 percent pass rate on its own eval set.
The author diffed pinned Python dependencies against model cutoffs and found stale calls in Claude models on small samples. Short notes in AGENTS.md raised the correct rate on held-out tasks. The author calls this a demonstration, not a benchmark.
The team reports Devin invented features and APIs, including a UI theme setting that did nothing, and struggled with less common tools such as nbdev and FastHTML.
Willison describes a model with search finding Google's migration docs on its own and porting his code to the newer Gen AI SDK, which shows the fix: let the model read current docs.