This article was originally published on avinashsangle.com.
Claude Managed Agents Outcomes is a public-beta feature, launched on May 6, 2026, that lets you hand the agent a rubric and have a separate grader model check every draft against it. If the grader returns needs_revision, the gaps flow back to the writer for another pass, up to max_iterations (default 3, max 20). Same hosted harness, no human in the loop.
TL;DR
Outcomes is a rubric-graded iteration loop built into the Managed Agents harness. You send one event, user.define_outcome, and the agent works until the grader says satisfied or hits max_iterations.
A separate grader (same model and tools as the writer, fresh context window) evaluates every draft. Its feedback is the only signal the writer gets back on each revision.
