Turn one-off Codex workflows into shared skills and plugins: OpenAI's FDE team says it cut ~90% of custom code

#What changed

In a post on OpenAI's site (Oct 22, 2026), Colin Jarvis, Head of Forward Deployed Engineering, says his team had first built custom harnesses for a large semiconductor customer's verification work (directed test generation, root-cause debugging, an architecture assistant). Those reached production in 3 to 6 months; OpenAI reports about 48% time saved on directed test generation and weeks of debugging compressed into hours. As Codex improved, one engineer rebuilt an app around it, and the team then moved the use cases onto a common harness with Codex at its core.

Custom apps became plugins and skills on common Codex primitives, and Codex itself became the interface, so the separate terminal UIs and CLIs were retired. OpenAI says roughly 90% of the custom code in each project went away. Rolling out to a new team became mostly a matter of swapping in that team's connectors or MCP servers and generating fresh evals with a dedicated skill. In one August 2026 case the first deployment took two months and the second team took one day.

#Who is affected

Teams that have built several custom AI tools or internal agents for different groups and now maintain separate UIs, CLIs, and glue code for each; consultants and platform teams rolling the same workflow out to multiple teams.

#What to do now

1. List the internal AI tools you maintain and mark which parts are real domain knowledge (evals, failure cases, procedures) versus glue code written around model limits.
2. Package the domain knowledge as Codex skills and plugins instead of standalone apps, and let Codex be the front end rather than a custom UI or CLI.
3. Keep team-specific access in connectors or MCP servers so the same skills can be reused by the next team.
4. Turn eval creation into a skill too, so each new team starts by generating its own evals.
5. Re-review glue code each time the model improves and delete what the model now handles.

#What is not confirmed

The numbers are OpenAI's own reporting about its own engagement, with an unnamed customer, and a 15-person FDE team behind it. Your savings will depend on how much of your current code is glue versus domain logic.

#Sources

• OpenAI, 'What FDE means at OpenAI' by Colin Jarvis: https://openai.com/index/forward-deployed-engineering-means
• Spotted via @Xudong07452910 on X: https://x.com/Xudong07452910/status/2113415101877559707