OpenAI uses Codex to build self-improving tax agents

Tax season is notoriously brutal for accountants. Entering data for a single medium-to-large return can easily eat up eight hours of manual digging through messy receipts, handwritten notes, and prior-year returns.

To tackle that bottleneck, OpenAI and Thrive Holdings teamed up to build Tax AI for Crete, a network of more than 30 accounting firms. Deployed during the busy season, the agent processed 7,000 returns across Form 1040 and Form 1041. It shaved roughly a third off tax preparation time, increased throughput by 50%, and drafted returns with up to 97% accuracy.

Yet the most remarkable part isn't the raw speedup. It is that the software spent the entire tax season teaching itself how to code better.

Turning red pen corrections into code fixes

When Tax AI first launched, it was barely passable. Only 25% of returns achieved 75% correct field completion. But within six weeks, that share rocketed to 86%, with even steeper gains at the 90% and 100% accuracy thresholds.

The system started with straightforward forms like W-2s and 1099s. As the weeks rolled on, it tackled far gnarlier filings involving Schedule E rental properties, K-1s, and reconciliations across messy multi-page documents.

Instead of engineers manually hunting down every extraction miss, the team turned practitioner reviews into an automated feedback loop. When a CPA corrected a value, the system logged the full trace: what Tax AI proposed, what the CPA changed, and the provenance tracing back to source files.

The pipeline grouped recurring mistakes—like repeatedly missing fair-rental-day fields—and automatically turned them into targeted eval suites.

Letting Codex debug the engine

Once a targeted eval was packaged, Codex took over the engineering triage.

Codex inspected the execution trace, repo files, and schemas to diagnose why a field failed. It pinpointed whether the bug was a missed extraction pattern, a flawed document selection step, or an unmapped field in the tax engine.

After diagnosing the root cause, Codex implemented the fix directly in the codebase, ran the new targeted eval alongside full regression suites, and opened a candidate pull request for human engineers to review.

Practitioners did their normal work, and their routine corrections directly fueled the agent's autonomous code upgrades.

Why it matters

We think this deployment outlines the real blueprint for enterprise AI agents: stop trying to make agents flawless in a vacuum, and start engineering feedback loops where real-world usage automatically improves the software.

For builders in high-liability verticals like accounting, legal, and healthcare, the takeaway is clear. Your moat isn't prompt engineering; it's eval infrastructure and structured production tracing. Turning human interventions into automated regression tests and pull requests shrinks product iteration from weeks to hours.

For the software industry, it marks a shift in how vertical software evolves. Instead of human developers slowly sifting through support tickets, Codex acts as an autonomous tier-one engineer maintaining its own pipeline.

That said, the approach requires deep discipline. Without airtight regression suites and rigorous human sign-off on generated PRs, autonomous self-improvement can easily veer into silent errors. But where high-stakes accuracy is strictly measured, this loop looks like the fastest way forward.