OpenAI engineers say that the internal proxy tool is the most fun project of their career.

SnapshotBot · 2026-03-28T08:25:01+00:00

OpenAI engineer Jason Liu shared his progress on the internal Codex proxy tool, stating that its usability has reached 99%. He has built 58 automation tools and 30 plugins, marking that the proxy system now has practical application value. It may be opened to the public API in the future to enhance enterprise automation efficiency.

SnapshotBot

2026-03-28 08:25:01

Abstract generation in progress

Title

OpenAI Engineer Says Internal Proxy Tool is the Most Fun Project of His Career

Summary

Jason Liu is a senior ML engineer at OpenAI, previously having worked at Stitch Fix and Meta. He recently posted that he is “having more fun than anyone at any AI lab.” The reason is that he has made new progress on OpenAI’s internal Codex proxy tool: the plugin just launched, achieving 99% usability. He has set up 58 automations and 30 plugins for his workflow.

This indicates that the internal proxy system has moved from the prototype stage to practical usability, having real significance for enterprise automation and production deployment.

Interpretation

Liu wrote the Instructor library, which OpenAI has publicly thanked for inspiring the approach of “structured LLM output.” Before joining OpenAI, he consulted for companies like Zapier and HubSpot. This focus on practical implementation makes his judgments more credible.

He stated that the system now reaches 99% usability, with automation covering most tasks except for “brainstorming and chatting with people.” This means OpenAI has made real progress in creating a reliable, scalable proxy system. He personally prefers the efficiency route of a single proxy rather than the experimental multi-proxy architecture.

Comparison of the two routes:

Dimension	Single Proxy (Liu’s Choice)	Multi-Proxy (Experimental)
Goal	End-to-end efficiency, simple operations	Division of labor, breaking down complex tasks
Risk	Fewer components, easier to troubleshoot	Complex coordination, faults can easily propagate
Current Status	Achieved 99% usability	Still exploring the best way to orchestrate

Competitive Perspective: If OpenAI’s internal tools approach production usability, these capabilities are likely to gradually open up to public APIs and developer tools.
Talent Perspective: Allowing engineers to “have fun” in real work is itself an attraction, helping to accelerate iteration.

Impact Assessment

Importance: High
Category: Technical Insight, Developer Tools, AI Research

Conclusion: This is still in the early stages, but signs of production usability are already visible. The first beneficiaries will be product developers and enterprise teams. There are currently no direct actionable signals for trading investors; long-term capital can focus on when internal capabilities will open up to public products and APIs.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.