OpenAI and Thrive Holdings detailed a collaboration to build Tax AI, a self-improving agent built on Codex that automates preparation of complex tax returns such as 1040 and 1041 forms. The system turns practitioner corrections and production traces into structured evaluation targets, surfacing repeated field-level errors as failure signals that Codex inspects alongside traces, evals, and skills to generate candidate fixes, while ambiguous cases route back to engineers for review. The collaboration with Thrive Holdings and Crete Professionals Alliance reports up to 97% accuracy and a roughly 50% increase in throughput, framed as a model for agentic self-improvement in other professional-services domains.
