What does FDE mean at OpenAI? About 90% less custom code with a Codex harness
Hi everyone, it's Shii-chan!
OpenAI News
openai.com
Today I read OpenAI's post on what Forward Deployed Engineering means for them. It's about making AI truly usable in real work, with lots of examples!
What was announced?
OpenAI News published an explainer on Forward Deployed Engineering (FDE), written by the person who founded the team in early 2025.
FDE aims to close the "deployment gap": the distance between what models seem capable of and what people can reliably use for their work. FDEs take on the hardest customer problems, turn the results into repeatable solutions, and feed what they learn back to Product and Research. Ideally, FDEs eventually become unnecessary for a given class of problem.
Why it matters
A flashy demo isn't enough if it can't hold up in production. The team's guiding principle is to start with the outcome, do what it takes to succeed, and only then scale.
What changes
The post shares how FDE work reached customers:
- Agriculture (John Deere): scale personalized advice for farmers. See & Spray can cut chemical use by up to 70%, but farmers still need help with setup and season planning. Using OpenAI APIs, the team built a natural-language assistant for tailored setup guidance, recommendations, underperformance alerts, and explanations of value
- Telecom (voice customer service): voice was about 90% of service volume and costlier than chat. Key production evals started around 60%, and with Research testing successive model snapshots, performance rose to 80% and then 90%+ over six months
- Life sciences: FDEs with scientific backgrounds earn researchers' trust on target validation, gene-expression analysis, and synthesis-route reconstruction
- Accounting (Thrive Holdings): a pilot processed 7,000 tax returns and raised throughput by about 50% with the same capacity
Dive Deep
Semiconductor verification
About 15 FDEs worked on 10+ use cases to make chip verification AI-native. Verification can account for roughly 70% of project effort in the industry.
- Built directed test generation, root cause analysis and debugging, and an architecture assistant
- Within 3-6 months of production: about 48% time saved on directed test generation
- 22% median performance improvement through kernel optimization
- Weeks of debugging compressed into hours
Making it repeatable
The first workflows were custom integrations that were hard to extend. The team moved them to a common harness with Codex at its core, replacing custom apps with plugins and skills built on shared Codex primitives. That removed roughly 90% of custom code per project.
- Codex improvements apply centrally, reducing glue code
- Skills and plugins are shareable and governed by the AI Center of Excellence
- Extending to a new team mostly means swapping connectors or MCP servers and building new evals
- The first deployment took two months; the second team took one day
OpenAI Presence, for enterprise voice and chat agents, grew out of real deployments that began with Swarm and evolved through the Agents API. One customer now handles more than half its calls with AI.
Two ways to make yourself obsolete
- Product improvement: a better product lets FDEs replace custom implementations and keep the value
- Model improvement: problems current models can't solve become targets for Research, as the telecom voice work shows
The team values enough domain knowledge to judge well, curiosity to challenge the status quo, and the simplest solution that works reliably. When Codex can handle what they built before, the answer is to go find the harder problem.
Wrap-up
- FDE closes the gap between model capability and real-world use
- It solves hard customer problems, makes them repeatable, and feeds back to Product and Research
- Semiconductor verification saw about 48% time saved on directed test generation; telecom evals rose from about 60% to 90%+
- A Codex-based shared harness removed about 90% of custom code and cut the second team's rollout to one day
This one is for anyone bringing AI into real company workflows, or helping others do it!
