agentic-eval

agentic-eval is an agent skill from the github/awesome-copilot collection that teaches iterative evaluation and refinement, so an agent can assess and improve its own output instead of accepting the first draft. The loop runs generate, evaluate, critique, refine and output, and it repeats until the result meets the criteria. The skill describes three core patterns. Basic reflection is a self-critique loop in which the agent reviews its own work. The evaluator-optimizer pattern separates generation from evaluation, so one step produces the output and another judges it, which tends to catch errors that the author misses. Code-specific test-driven refinement uses tests as the evaluator, so changes are kept only when they pass. The skill says when the approach pays off: quality-critical generation such as code, reports and analysis that must be accurate, tasks with clear evaluation criteria where success can be defined, and content that must meet a style guide, compliance rules or formatting requirements. It is less useful when there is no way to judge the output, because a loop with no criteria just rewrites text. The patterns cost extra model calls, so the number of rounds should be bounded. The skill is installed with the skills CLI from the github/awesome-copilot repository. Its skills.sh page showed about 10.2K installs and passing results from three security audits at the time of the scan.…