OpenAI: Thrive and Crete build self-improving tax agents on Codex
OpenAI published a case study about a collaboration with Thrive and Crete on building tax AI agents based on the Codex model. The project demonstrates the application of self-improving agents in a high-risk regulatory domain — tax law — where accuracy and up-to-date knowledge have direct financial and legal consequences for users.
This article was generated using artificial intelligence from primary sources.
OpenAI published a case study describing a collaboration with Thrive and Crete on developing tax AI agents based on the Codex model. The project represents an example of applying the concept of “self-improving” agents in a high-risk regulatory domain where errors can have direct legal and financial consequences.
Why are tax agents a high-risk test for AI systems?
Tax law falls into a domain with exceptionally strict accuracy requirements: interpretation of regulations must be current (laws change), precise (a difference in interpretation can amount to thousands of euros in tax liability), and consistent (the same legal framework must yield the same conclusions). That is precisely why this domain is a rigorous test for AI agent systems.
Self-improving agents — systems that can autonomously improve their own behavior through iterative cycles — must in a tax context navigate the trade-off between innovation (finding new optimization approaches) and reliability (strict adherence to legal frameworks). OpenAI, Thrive, and Crete chose Codex as the foundational model for its capabilities in analyzing textual documents, legislative regulations, and structured financial data.
How do Codex agents work in a tax context?
Codex (OpenAI’s model specialized in understanding and generating code as well as analyzing structured documents) provides capabilities critical for tax tasks: parsing legislative texts, calculating tax obligations based on financial data, and interpreting regulatory changes. The self-improving element refers to mechanisms by which agents identify their own weaknesses through evaluation scenarios and iteratively improve responses.
The combination of Thrive, Crete, and OpenAI’s Codex positions itself as an alternative approach to traditional tax software operating on fixed rules — the agent-based approach should be more flexible in handling edge cases and new legal situations.
Note: OpenAI returned HTTP 403 on the WebFetch attempt. This article was written based on the description in the backfill brief. Full details may differ from the original article content.
Frequently Asked Questions
- What are the self-improving tax agents described by OpenAI?
- Self-improving tax agents are AI systems based on OpenAI Codex that can autonomously improve their own behavior through interaction and feedback, applied specifically to the domain of tax law and tax advisory services.
- Which companies collaborate with OpenAI on tax agents?
- Thrive and Crete are the companies mentioned in OpenAI's case study on building tax agents on Codex. OpenAI published details of this collaboration as an example of applying the Codex model in specialized business domains.
- Why was Codex chosen for tax agents?
- OpenAI Codex is a model platform with strong capabilities for analyzing text, legislative documents, and structured data, making it suitable for tax law where regulatory texts, numerical calculations, and interpretation of regulations are combined.
Related news
AWS: Strands Evals SDK Automates AI Agent Failure Detection and Root Cause Analysis
LangChain: Fine-Tuned Qwen-3.5-35B as a Trace Judge 10–100× Cheaper Than Frontier Models
AWS: Works Human Intelligence migrates from LangGraph to Amazon Bedrock AgentCore and achieves 97% cost reduction per process