We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·May 27·5 min read
Building self-improving tax agents with Codex
TL;DR
Tax filing involves manual data entry and high error rates, creating bottlenecks in accounting workflows. OpenAI, Thrive, and Crete built a self-improving tax agent using Codex (a code-generating AI model) that automates filing tasks and learns from corrections.
✦ Why It Matters
Engineers can apply self-improving code-generation patterns to automate complex, rule-based workflows in regulated domains.
Key Takeaways
- Use code-generating models like Codex to automate rule-heavy, document-processing tasks.
- Build feedback mechanisms so agents learn from user corrections and improve iteratively.
- Measure impact on manual effort reduction and error rates in regulated workflows.
- Apply this pattern to other domains requiring document interpretation and automated compliance.
Full Summary
Tax preparation traditionally requires manual data extraction, form completion, and verification—processes prone to human error and time-consuming. OpenAI partnered with Thrive and Crete to develop a self-improving tax agent powered by Codex, a large language model trained to generate and understand code.
The agent uses Codex to interpret tax documents, generate filing logic, and automate submission workflows. Critically, the system incorporates feedback loops: when users correct errors or provide clarifications, the agent learns and refines its approach for future filings.
Results included measurable reductions in manual intervention time, higher accuracy rates on complex returns, and faster end-to-end processing. This approach demonstrates how code-generating models can tackle domain-specific automation beyond traditional software engineering.