Cult.fit expands OpenAI Codex use across its fitness business

Codex becomes a shared development platform

Cult.fit operates 708 fitness centers across 77 cities and serves nearly one million paid members. Its business includes gym subscriptions, trainer-led classes, fitness equipment, accessories, workout footwear and activewear.

Before adopting Codex as its standard platform, cult.fit developers used a mix of coding agents and assistants. The company said that made it harder to share knowledge and reuse capabilities because teams followed different conventions for repositories, agent instructions, skills and development workflows.

Codex now provides common conventions for repository structures, testing, code review and feedback loops. Cult.fit has also built specialized workflows, including a Site Reliability Engineering agent that can aggregate observability data and logs, investigate incidents and, in some cases, create a pull request with a proposed fix for an engineer to review.

Security agents help identify and address vulnerabilities, while other on-call agents investigate recurring production issues. The company is also moving toward spec-driven development, where a specification guides work through the software development lifecycle toward a reviewable pull request.

According to cult.fit CTO Santhosh Kumar, a social feed for the cult.fit app that would previously have taken more than a year was built and shipped end to end in two months using the new workflow.

Non-engineering teams build their own tools

Cult.fit has expanded Codex into other parts of the organization by making company data easier for teams to access and analyze. Previously, even small data requests had to go through a central analyst team. The company says more than 20 internal applications have been built by non-engineering teams in recent months.

Its HR team created a Codex workflow for performance reviews, scheduling, feedback collection and consolidating the resulting information. In another example, an engineer rebuilt an internal dashboard capability over a weekend after the cost of the company’s software increased sharply. Cult.fit says running the replacement on its own application avoids more than $100,000 in annual software costs.

The rollout included a company hackathon, a one-day session in which management leaders built tools for their own functions, and twice-weekly engineering office hours. Engineers are also embedded in selected functions to provide technical and domain support.

Cult.fit reports a reduction of more than 50% in customer support costs after adding AI agents to its existing Sprinklr support platform and connecting them to backend systems. One workflow assesses requests to waive penalties for missed group fitness classes by checking signals such as local weather and attendance patterns before deciding whether a penalty should be waived.

AI Trainer adds voice-based coaching

Cult.fit is using the OpenAI API to update its AI Trainer, which was created to provide guided, personalized workouts without a personal trainer. More than 45% of cult.fit members are first-time gym goers, according to the company, and around 20% of its gym users currently use the AI Trainer.

The latest version is designed around a voice-first workout experience. Through an OpenAI API integration using GPT Realtime, it can provide coaching cues and support workout logging through two-way voice interaction.

Cult.fit’s stated next step is to move from individual AI workflows toward an operating model in which AI is built into how work gets done. Kumar describes the goal for operations as an “autopilot operations org,” while the member experience ambition is a “trainer in your pocket.”