The Pulse
OpenAI Says Codex Should Make Its Field Engineers Obsolete
OpenAI says its forward-deployed engineers work with customers to move AI systems into production, then turn successful projects into reusable tools. Head of Forward Deployed Engineering Colin Jarvis cites deployments in semiconductor desig
AI.info Team · Written with AI · fact-checked ·
OpenAI says its Codex tools should eventually make some of the customer-specific systems built by its forward-deployed engineers unnecessary. That prospect could seem to undermine the team’s role; Colin Jarvis, who leads it, presents it as a measure of success. In a post published October 22, 2026, Jarvis describes a cycle of working directly with customers, adapting AI to difficult production tasks, and turning what works into reusable software.
“If we make ourselves obsolete every couple of months, our FDE team is doing its job.”
Colin Jarvis, Head of Forward Deployed Engineering at OpenAI
From customer problems to reusable systems
Jarvis says he founded OpenAI’s Forward Deployed Engineering, or FDE, team in early 2025 to close what the company calls the “deployment gap”: the distance between what models appear able to do and what people can reliably use at work. Engineers work alongside customers to understand operational constraints, build solutions for problems existing products cannot handle, and share the resulting lessons with OpenAI’s Product and Research teams.
The approach has two paths. When a model can already do the job, FDE works to make the solution repeatable; when it falls short, the team helps define the gap and works with Research on improvements. Jarvis says the goal is to leave the next customer with a better starting point, rather than a permanent dependence on bespoke engineering.
A semiconductor project puts the numbers on it
Jarvis describes a mid-2025 effort with a large, unnamed semiconductor customer. Around 15 FDEs worked across more than 10 use cases, including test generation, debugging and an architecture assistant. OpenAI’s post says verification can account for roughly 70% of project effort in the industry.
The engineers built custom evaluation tools with the customer’s verification staff and brought the applications into production over three to six months. OpenAI reports about 48% time saved through AI-based directed test generation, a 22% median performance improvement through kernel optimization, and debugging work reduced from weeks to hours. Jarvis says some team members had joined only days earlier and had no semiconductor experience; they learned the domain while working with the customer’s engineers.
Codex cuts the custom code
Once those workflows showed value, the team faced a different problem: each required too much custom integration to spread easily. Jarvis says an FDE rebuilt the applications around Codex, moving them onto a shared harness and replacing separate terminal interfaces with Codex-based plugins and skills.
OpenAI says that change removed roughly 90% of the custom code from each project. In an example from August 2026, the initial deployment took two months, while bringing the solution to a second team took one day. Jarvis frames that shift as a practical test of the team’s purpose: prove a workflow in demanding conditions, then reduce the work needed to deploy it again.
Model limits become research questions
Not every deployment can be solved with better integration. For a large, unnamed telecommunications customer, voice made up around 90% of customer-service volume, but the model scored about 60% on the evaluations used to judge production readiness. The work involved challenges including accents, tone, latency and accuracy.
FDE worked with the customer to build the platform and specify what the model needed to do; Research worked on the underlying capabilities. OpenAI says performance on the key evaluations rose from around 60% to 80%, then above 90%, over six months. The account makes the team’s trade-off explicit: its engineers take on difficult, sometimes one-off customer work, then try to turn the lessons into product improvements, more capable models or less custom work for whoever comes next.