Codex agents that self-improve in production

OpenAI and Thrive Holdings built Tax AI, a system that uses Codex to prepare complex tax returns for over 30 accounting firms. The key innovation isn’t just automation—it’s that the agent learns from real-world use without waiting for engineers to manually fix every edge case. In six weeks, the system jumped from 25% to 86% accuracy at the 75% field completion mark. Tax AI handles messy documents, extracts data, and creates ready-to-review submissions, saving accountants about a third of their time and boosting throughput by 50%. The secret sauce is an eval infrastructure that turns production feedback into structured signals for Codex to self-improve. Why it matters: This approach points toward a future where AI systems don’t just get smarter between deployments—they get smarter every time someone uses them.