Build trace-driven agent improvement loop with Promptfoo evals, HALO-ranked harness changes, and Codex handoff

AI Impact Summary

This initiative creates a trace-driven feedback loop that converts human and model interactions into structured Promptfoo evals, HALO-ranked harness adjustments, and Codex handoffs. By codifying traces into evaluative data and harness changes, teams can systematically improve agent reliability and code-generation alignment, reducing trial-and-error in future iterations. The effort touches instrumentation, data governance, and CI pipelines to support continuous evaluation and safe handoffs to Codex when code tasks are involved.