Anthropic's 'Outcomes' Lifted Quality 10% — By Adding a Grader, Not a Bigger Model

Anthropic shipped a feature at Code with Claude 2026 that should reframe how you think about output quality. On its own benchmarks, "Outcomes" raised generated-document quality by 8.4% for Word and 10.1% for PowerPoint — with no model upgrade at all. The entire gain came from one structural change: a second agent that grades the first agent's output against a rubric you write.

The mechanism is boring, which is the point. Your task agent finishes its work and hands the result to a grading agent. That grader reads your rubric — your definition of "good" — scores the output, and kicks it back for another pass if it falls below threshold. No new weights, no bigger context window. Just a verification loop bolted onto the same model.

This is the lesson prototypes never learn. A demo proves the model can draft a deck. Production quality comes from the layer that decides whether the draft is actually good — the rubric, the score, the retry. That layer is code you own, not a checkpoint you wait for.

We watch teams chase a 3-point benchmark bump on the next release while a well-written rubric and a grading pass sit on the table, worth 10. The moat isn't the model. It's the spec for "done."

See how we build the eval layer →