1 min lesson
One agent run, measured
Tell someone how to act on this idea: "You can list at least three things the harness does that a raw API call does not."
Step 1 of 3
The case: you can name what the harness adds over a raw API call and say why the same model runs better and cheaper through it. Start with this check: You can list at least three things the harness does that a raw API call does not. Walk it once against a real situation so you can defend the call, not just assert it.
Learn more
Full explanation
One agent run, measured
Harness vs raw API — show why the same model runs cheaper through Cursor
One agent run, measured
SaySomeone on the team says Cursor is just a wrapper around a model API key, so bringing your own key would be cheaper. Walk one real task through the harness and show what a raw call skips.
DoRun a normal agent task and watch dynamic context discoveryThe agent pulling only the relevant parts of files, tools and MCP servers into context as needed, instead of loading everything up front. Press Enter for the full definition.: instead of loading the whole repo, the harness pulls only the relevant files, snippets and tool results into the window.
SeeThe request that reaches the model is smaller and better-shaped than a raw call that dumps everything in.
DoNote the two efficiency layers a raw call has to build itself: caching that reuses earlier context, and retries that recover from transient failures instead of re-spending tokens.
SayOn Cursor's own measure, caching plus dynamic context discoveryThe agent pulling only the relevant parts of files, tools and MCP servers into context as needed, instead of loading everything up front. Press Enter for the full definition. cut agent tokens by around 47%, and that holds even when the model behind it belongs to another provider.
SeeFewer tokens for the same work, so going through the harness can beat bringing your own key on cost.
SeeThe proof point: you can name at least three things the harness adds over a raw call - context selection, caching and per-model tuning - and say why the same model runs better and cheaper through Cursor than direct. The model is the last layer, not the only one.
Learn more
Optional practice
Practice: One agent run, measured
QYou've applied it. What actually proves the work is done, not just a plausible answer?