Building a Self-Improving Tax Agent with Codex

Systems in real-world production environments, unlike labs, produce errors in unpredictable ways. Previously, engineers had to manually analyze feedback and modify prompts, but combining Codex with sophisticated evaluation infrastructure makes it possible to build a self-improving system where the agent improves its own performance.

OpenAI and Thrive Holdings partnered with over 30 accounting firms at Crete to develop Tax AI. This system uses Codex to convert real-world usage data into structured signals, driving autonomous performance improvement.

Tax AI processed 7,000 tax filings this season, automating the preparation of 1040 and 1041 returns. As a result, it achieved the following outcomes:

• Reduced tax preparation time by about 1/3
• Achieved accuracy of up to 97%
• Increased throughput by about 50%

Accuracy also showed dramatic growth. At launch, only 25% of returns had a field completion rate above 75%, but within six weeks this surged to 86%.

This self-improvement is driven by three key pillars:

• Expert practitioner feedback
• Production traces: structured records from input to final output
• Codex-driven iteration loop: continuous development through custom evaluations