Creating Codex-powered tax agents that improve themselves

Meta description: Discover how OpenAI, Thrive Capital, and Crete Professionals built a self-improving tax agent with Codex to automate filings and boost accuracy.

Tags: OpenAI, Codex, AI agents, tax automation, Thrive Capital

Featured image: Abstract, friendly illustration of a glowing loop of code and documents flowing through a stylized AI brain, soft blue and orange palette, no people or logos.

How OpenAI, Thrive, and Crete Built a Self-Improving Tax Agent

If you've ever wondered whether AI can handle genuinely complicated professional work—not just drafting emails, but real, high-stakes tasks like tax filings—a new project from OpenAI, Thrive Capital, and Crete Professionals is worth a close look. The three organizations teamed up to build an AI tax agent powered by Codex, OpenAI's coding-focused model, and the result is a system that prepares tax filings, checks its own work, and gets better over time.

In this article, we'll walk through who was involved, how the system actually works, and what the collaboration suggests about the future of AI in professional services. Whether you're a founder exploring automation, an accountant curious about AI, or just someone tracking where AI agents are headed, here's the plain-English breakdown.

The Team Behind the Project

The collaboration brings together three very different players, and each one fills a specific role.

OpenAI contributes the core technology: Codex, its model designed for writing and refining code. Tax preparation might not sound like a coding problem, but it turns out that much of the work—pulling data from documents, applying tax rules, generating filing outputs—can be expressed as structured, repeatable tasks that a coding model handles well.

Thrive Capital acts as the connective tissue. The venture firm has been active in pairing AI companies with real businesses, and in this project it helped orchestrate the partnership and push the work toward practical deployment rather than staying stuck in the lab.

Crete Professionals supplies the domain expertise. Crete operates an AI-enabled platform for accounting and tax work, so its team knows exactly where the bottlenecks are in real filings: the repetitive data entry, the cross-checking, the parts of the job that eat up skilled professionals' time without requiring their judgment.

The idea is simple in hindsight: a frontier AI lab with the model, a firm with the workflows, and an investor with the incentive to make both click together.

What the Tax Agent Actually Does

The agent is built to take on meaningful chunks of the tax preparation process, not just assist around the edges. According to reporting on the project, the system automates significant portions of filing work—gathering and organizing the information needed for returns, applying the relevant rules, and producing outputs that human professionals can then review.

The phrase that keeps coming up in coverage is "self-improving." That's the part that makes this project notable. Rather than a static tool that performs the same way on day 100 as it did on day one, the agent is designed to learn from its work: when reviewers catch issues or corrections, those lessons feed back into how the agent handles similar tasks next time. It's a loop—do the work, get feedback, improve, repeat.

Accuracy is the other headline claim. Tax work is unforgiving; errors carry real consequences. So the emphasis isn't just speed, but building a system whose output becomes more reliable with each cycle. Human professionals remain in the picture as reviewers and decision-makers, which is a sensible setup for now—AI does the heavy lifting, people supply judgment and accountability.

Why Codex Instead of a Chatbot

One of the more interesting choices here is the use of Codex rather than a general-purpose conversational model. Tax preparation is full of structured, rule-driven processes: specific forms, specific calculations, specific validation steps. Those are exactly the kinds of tasks that benefit from an agentic system that can execute multi-step workflows, run checks, and produce consistent outputs—rather than simply answering questions in a chat window.

By framing tax work as something closer to a software pipeline, the team turned a messy, judgment-heavy document problem into something a model can tackle step by step, with verification built in at each stage. It's a useful lesson for anyone building AI tools for professional services: the winning approach usually isn't "replace the expert," it's "automate the structured layers underneath the expert."

What This Says About AI in Professional Work

This project is one more data point in a bigger shift. For years, the question was whether AI could handle knowledge work at all. Now the question is becoming which parts of that work can be safely handed off, and how quickly the systems improve once they are.

The self-improving angle matters here. A tool that gets better through real use creates a compounding advantage—early adopters don't just save time today, they accumulate a system that's increasingly tuned to their exact workflows. For accounting firms and similar businesses, that reframes the build-vs-buy conversation entirely.

At the same time, it's worth staying level-headed. Tax work involves legal and regulatory nuance, and the professionals reviewing the agent's output aren't going away anytime soon. The realistic near-term picture is human-AI collaboration: machines handling volume and repetition, humans handling edge cases and final sign-off.

The Takeaway

The OpenAI, Thrive, and Crete collaboration is a concrete example of AI agents moving from demos into real professional workflows. A coding model, a domain expert, and an investor aligned around a shared goal produced a system that automates filings, improves with feedback, and speeds up work that used to drain hours from skilled professionals.

If you're watching where agentic AI goes next, watch projects like this one. The pattern—specialized model, real workflow, feedback loop—is likely to repeat across law, finance, healthcare administration, and beyond. The tools are getting better not just because the models improve, but because teams are learning how to wrap them in processes that make improvement automatic.

The bottom line: self-improving AI agents are no longer a concept on a slide deck. They're preparing actual tax filings, and they're getting sharper at it with every cycle. That should get the attention of anyone whose job involves structured professional work—and anyone building the tools that will change it.