Building self-improving tax agents with Codex

Source: OpenAI Date: 2026-05-27 URL: https://openai.com/index/building-self-improving-tax-agents-with-codex

Summary

OpenAI published a case study describing tax agents built on Codex that improve their own performance over time — a demonstration of agents that can generate their own training signal, evaluate outputs, and iteratively refine their approach without human-authored feedback at each step. The source page returned 403; this summary is grounded in the title and the broader Codex-in-professional-domains context from the same period. The specific mechanism of self-improvement (synthetic QA, RL fine-tuning, tool-use traces, or another method) is not confirmed.

Implications

• Agentic engineering patterns. Self-improving agents in high-stakes professional domains (tax, legal, finance) represent a qualitative shift from task-execution agents to agents that model their own performance. The “self-improving” label is doing real work here: if Codex agents can generate and validate their own feedback loops in structured domains, the human oversight requirement shifts from output review to oversight of the improvement process itself.
• Vendor/lab strategy. OpenAI publishing professional-domain agent case studies in the same window as Codex CLI expansion signals a deliberate enterprise narrative: Codex is not just a coding assistant but a professional-agent substrate. This positions it against Anthropic’s Managed Agents and Claude for Work directly.
• Model landscape. Self-improvement via agent traces (rather than human preference data) is a capability story that matters for fine-tuning economics — if agents can generate their own high-quality training signal in specialized domains, the cost of domain adaptation drops substantially.