Outcome Based Sessions

"Outcome Based Sessions" is the new, operationalized evolution of the Ralph Wiggum pattern. Amazing to watch an idea move through an industry in real-time.

https://platform.claude.com/docs/en/managed-agents/define-outcomes

The operationalization evolves the pattern into common language:

• outcome — the prompt/spec that describes the end state
• grader — an auto-eval that uses LLM as a judge using a rubric, ran each loop
• rubric — a prompt/skill that defines validation/eval criteria
• max-iterations — same feature as in the Ralph Wiggum pattern
• interrupt — a user injecting additional steering into a running session
• outcome_eval event — surfaces an ongoing eval score on iteration complete